Category: Offsites

Posted by Dana Alon and Jeongwoo Ko, Software Engineers, Google Research

Emotions are a key aspect of social interactions, influencing the way people behave and shaping relationships. This is especially true with language — with only a few words, we’re able to express a wide variety of subtle and complex emotions. As such, it’s been a long-term goal among the research community to enable machines to understand context and emotion, which would, in turn, enable a variety of applications, including empathetic chatbots, models to detect harmful online behavior, and improved customer support interactions.

In the past decade, the NLP research community has made available several datasets for language-based emotion classification. The majority of those are constructed manually and cover targeted domains (news headlines, movie subtitles, and even fairy tales) but tend to be relatively small, or focus only on the six basic emotions (anger, surprise, disgust, joy, fear, and sadness) that were proposed in 1992. While these emotion datasets enabled initial explorations into emotion classification, they also highlighted the need for a large-scale dataset over a more extensive set of emotions that could facilitate a broader scope of future potential applications.

In “GoEmotions: A Dataset of Fine-Grained Emotions”, we describe GoEmotions, a human-annotated dataset of 58k Reddit comments extracted from popular English-language subreddits and labeled with 27 emotion categories . As the largest fully annotated English language fine-grained emotion dataset to date, we designed the GoEmotions taxonomy with both psychology and data applicability in mind. In contrast to the basic six emotions, which include only one positive emotion (joy), our taxonomy includes 12 positive, 11 negative, 4 ambiguous emotion categories and 1 “neutral”, making it widely suitable for conversation understanding tasks that require a subtle differentiation between emotion expressions.

We are releasing the GoEmotions dataset along with a detailed tutorial that demonstrates the process of training a neural model architecture (available on TensorFlow Model Garden) using GoEmotions and applying it for the task of suggesting emojis based on conversational text. In the GoEmotions Model Card we explore additional uses for models built with GoEmotions, as well as considerations and limitations for using the data.

This text expresses several emotions at once, including excitement, approval and gratitude.

This text expresses relief, a complex emotion conveying both positive and negative sentiment.

This text conveys remorse, a complex emotion that is expressed frequently but is not captured by simple models of emotion.

Building the Dataset
Our goal was to build a large dataset, focused on conversational data, where emotion is a critical component of the communication. Because the Reddit platform offers a large, publicly available volume of content that includes direct user-to-user conversation, it is a valuable resource for emotion analysis. So, we built GoEmotions using Reddit comments from 2005 (the start of Reddit) to January 2019, sourced from subreddits with at least 10k comments, excluding deleted and non-English comments.

To enable building broadly representative emotion models, we applied data curation measures to ensure the dataset does not reinforce general, nor emotion-specific, language biases. This was particularly important because Reddit has a known demographic bias leaning towards young male users, which is not reflective of a globally diverse population. The platform also introduces a skew towards toxic, offensive language. To address these concerns, we identified harmful comments using predefined terms for offensive/adult and vulgar content, and for identity and religion, which we used for data filtering and masking. We additionally filtered the data to reduce profanity, limit text length, and balance for represented emotions and sentiments. To avoid over-representation of popular subreddits and to ensure the comments also reflect less active subreddits, we also balanced the data among subreddit communities.

We created a taxonomy seeking to jointly maximize three objectives: (1) provide the greatest coverage of the emotions expressed in Reddit data; (2) provide the greatest coverage of types of emotional expressions; and (3) limit the overall number of emotions and their overlap. Such a taxonomy allows data-driven fine-grained emotion understanding, while also addressing potential data sparsity for some emotions.

Establishing the taxonomy was an iterative process to define and refine the emotion label categories. During the data labeling stages, we considered a total of 56 emotion categories. From this sample, we identified and removed emotions that were scarcely selected by raters, had low interrater agreement due to similarity to other emotions, or were difficult to detect from text. We also added emotions that were frequently suggested by raters and were well represented in the data. Finally, we refined emotion category names to maximize interpretability, leading to high interrater agreement, with 94% of examples having at least two raters agreeing on at least 1 emotion label.

The published GoEmotions dataset includes the taxonomy presented below, and was fully collected through a final round of data labeling where both the taxonomy and rating standards were pre-defined and fixed.

GoEmotions taxonomy: Includes 28 emotion categories, including “neutral”.

<!–

Positive		Negative		Ambiguous
admiration 👏	joy 😃	anger 😡	grief 😢	confusion 😕
amusement 😂	love ❤️	annoyance 😒	nervousness 😬	curiosity 🤔
approval 👍	optimism 🤞	disappointment	remorse 😔	realization 💡
caring 🤗	pride 😌	disapproval 👎	sadness 😞	surprise 😲
desire 😍	relief 😅	disgust 🤮
excitement 🤩		embarrassment 😳
gratitude 🙏		fear 😨

–>

Data Analysis and Results
Emotions are not distributed uniformly in the GoEmotions dataset. Importantly, the high frequency of positive emotions reinforces our motivation for a more diverse emotion taxonomy than that offered by the canonical six basic emotions.

To validate that our taxonomic choices match the underlying data, we conduct principal preserved component analysis (PPCA), a method used to compare two datasets by extracting linear combinations of emotion judgments that exhibit the highest joint variability across two sets of raters. It therefore helps us uncover dimensions of emotion that have high agreement across raters. PPCA was used before to understand principal dimensions of emotion recognition in video and speech, and we use it here to understand the principal dimensions of emotion in text.

We find that each component is significant (with p-values < 1.5e-6 for all dimensions), indicating that each emotion captures a unique part of the data. This is not trivial, since in previous work on emotion recognition in speech, only 12 out of 30 dimensions of emotion were found to be significant.

We examine the clustering of the defined emotions based on correlations among rater judgments. With this approach, two emotions will cluster together when they are frequently co-selected by raters. We find that emotions that are related in terms of their sentiment (negative, positive and ambiguous) cluster together, despite no predefined notion of sentiment in our taxonomy, indicating the quality and consistency of the ratings. For example, if one rater chose “excitement” as a label for a given comment, it is more likely that another rater would choose a correlated sentiment, such as “joy”, rather than, say, “fear”. Perhaps surprisingly, all ambiguous emotions clustered together, and they clustered more closely with positive emotions.

Similarly, emotions that are related in terms of their intensity, such as joy and excitement, nervousness and fear, sadness and grief, annoyance and anger, are also closely correlated.

Our paper provides additional analyses and modeling experiments using GoEmotions.

Future Work: Alternatives to Human-Labeling
While GoEmotions offers a large set of human-annotated emotion data, additional emotion datasets exist that use heuristics for automatic weak-labeling. The dominant heuristic uses emotion-related Twitter tags as emotion categories, which allows one to inexpensively generate large datasets. But this approach is limited for multiple reasons: the language used on Twitter is demonstrably different than many other language domains, thus limiting the applicability of the data; tags are human generated, and, when used directly, are prone to duplication, overlap, and other taxonomic inconsistencies; and the specificity of this approach to Twitter limits its applications to other language corpora.

We propose an alternative, and more easily available heuristic in which emojis embedded in user conversation serve as a proxy for emotion categories. This approach can be applied to any language corpora containing a reasonable occurence of emojis, including many that are conversational. Because emojis are more standardized and less sparse than Twitter-tags, they present fewer inconsistencies.

Note that both of the proposed approaches — using Twitter tags and using emojis — are not directly aimed at emotion understanding, but rather at variants of conversational expression. For example, in the conversation below, 🙏 conveys gratitude, 🎂 conveys a celebratory expression, and 🎁 is a literal replacement for ‘present’. Similarly, while many emojis are associated with emotion-related expressions, emotions are subtle and multi-faceted, and in many cases no one emoji can truly capture the full complexity of an emotion. Moreover, emojis capture varying expressions beyond emotions. For these reasons, we consider them as expressions rather than emotions.

This type of data can be valuable for building expressive conversational agents, as well as for suggesting contextual emojis, and is a particularly interesting area of future work.

Conclusion
The GoEmotions dataset provides a large, manually annotated, dataset for fine-grained emotion prediction. Our analysis demonstrates the reliability of the annotations and high coverage of the emotions expressed in Reddit comments. We hope that GoEmotions will be a valuable resource to language-based emotion researchers, and will allow practitioners to build creative emotion-driven applications, addressing a wide range of user emotions.

Acknowledgements
This blog presents research done by Dora Demszky (while interning at Google), Dana Alon (previously Movshovitz-Attias), Jeongwoo Ko, Alan Cowen, Gaurav Nemade, and Sujith Ravi. We thank Peter Young for his infrastructure and open sourcing contributions. We thank Erik Vee, Ravi Kumar, Andrew Tomkins, Patrick Mcgregor, and the Learn2Compress team for support and sponsorship of this research project.

Posted by Artem Dementyev, Hardware Engineer, Google Research

Most wearable smart devices and mobile phones have the means to communicate with the user through tactile feedback, enabling applications from simple notifications to sensory substitution for accessibility. Typically, they accomplish this using vibrotactile actuators, which are small electric vibration motors. However, designing a haptic system that is well-targeted and effective for a given task requires experimentation with the number of actuators and their locations in the device, yet most practical applications require standalone on-body devices and integration into small form factors. This combination of factors can be difficult to address outside of a laboratory as integrating these systems can be quite time-consuming and often requires a high level of expertise.

A typical lab setup on the left and the VHP board on the right.

In “VHP: Vibrotactile Haptics Platform for On-body Applications”, presented at ACM UIST 2021, we develop a low-power miniature electronics board that can drive up to 12 independent channels of haptic signals with arbitrary waveforms. The VHP electronics board can be battery-powered, and integrated into wearable devices and small gadgets. It allows all-day wear, has low latency, battery life between 3 and 25 hours, and can run 12 actuators simultaneously. We show that VHP can be used in bracelet, sleeve, and phone-case form factors. The bracelet was programmed with an audio-to-tactile interface to aid lipreading and remained functional when worn for multiple months by developers. To facilitate greater progress in the field of wearable multi-channel haptics with the necessary tools for their design, implementation, and experimentation, we are releasing the hardware design and software for the VHP system via GitHub.

Front and back sides of the VHP circuit board.

Block diagram of the system.

Platform Specifications.
VHP consists of a custom designed circuit board, where the main components are the microcontroller and haptic amplifier, which converts microcontroller’s digital output into signals that drive the actuators. The haptic actuators can be controlled by signals arriving via serial, USB, and Bluetooth Low Energy (BLE), as well as onboard microphones, using an nRF52840 microcontroller, which was chosen because it offers many input and output options and BLE, all in a small package. We added several sensors into the board to provide more experimental flexibility: an on-board digital microphone, an analog microphone amplifier, and an accelerometer. The firmware is a portable C/C++ library that works in the Arduino ecosystem.

To allow for rapid iteration during development, the interface between the board and actuators is critical. The 12 tactile signals’ wiring have to be quick to set up in order to allow for such development, while being flexible and robust to stand up to prolonged use. For the interface, we use a 24-pin FPC (flexible printed circuit) connector on the VHP. We support interfacing to the actuators in two ways: with a custom flexible circuit board and with a rigid breakout board.

VHP board (small board on the right) connected to three different types of tactile actuators via rigid breakout board (large board on the left).

Using Haptic Actuators as Sensors
In our previous blog post, we explored how back-EMF in a haptic actuator could be used for sensing and demonstrated a variety of useful applications. Instead of using back-EMF sensing in the VHP system, we measure the electrical current that drives each vibrotactile actuator and use the current load as the sensing mechanism. Unlike back-EMF sensing, this current-sensing approach allows simultaneous sensing and actuation, while minimizing the additional space needed on the board.

One challenge with the current-sensing approach is that there is a wide variety of vibrotactile actuators, each of which may behave differently and need different presets. In addition, because different actuators can be added and removed during prototyping with the adapter board, it would be useful if the VHP were able to identify the actuator automatically. This would improve the speed of prototyping and make the system more novice-friendly.

To explore this possibility, we collected current-load data from three off-the-shelf haptic actuators and trained a simple support vector machine classifier to recognize the difference in the signal pattern between actuators. The test accuracy was 100% for classifying the three actuators, indicating that each actuator has a very distinct response.

Different actuators have a different current signature during a frequency sweep, thus allowing for automatic identification.

Additionally, vibrotactile actuators require proper contact with the skin for consistent control over stimulation. Thus, the device should measure skin contact and either provide an alert or self-adjust if it is not loaded correctly. To test whether a skin contact measuring technique works in practice, we measured the current load on actuators in a bracelet as it was tightened and loosened around the wrist. As the bracelet strap is tightened, the contact pressure between the skin and the actuator increases and the current required to drive the actuator signal increases commensurately.

Current load sensing is responding to touch, while the actuator is driven at 250 Hz frequency.

Quality of the fit of the bracelet is measured.

Audio-to-Tactile Feedback
To demonstrate the utility of the VHP platform, we used it to develop an audio-to-tactile feedback device to help with lipreading. Lipreading can be difficult for many speech sounds that look similar (visemes), such as “pin” and “min”. In order to help the user differentiate visemes like these, we attach a microphone to the VHP system, which can then pick up the speech sounds and translate the audio to vibrations on the wrist. For audio-to-tactile translation, we used our previously developed algorithms for real-time audio-to-tactile conversion, available via GitHub. Briefly, audio filters are paired with neural networks to recognize certain viesemes (e.g., picking up the hard consonant “p” in “pin”), and are then translated to vibrations in different parts of the bracelet. Our approach is inspired by tactile phonemic sleeve (TAPS), however the major difference is that in our approach the tactile signal is presented continuously and in real-time.

One of the developers who employs lipreading in daily life wore the bracelet daily for several months and found it to give better information to facilitate lipreading than previous devices, allowing improved understanding of lipreading visemes with the bracelet versus lipreading alone. In the future, we plan to conduct full-scale experiments with multiple users wearing the device for an extended time.

Left: Audio-to-tactile sleeve. Middle: Audio-to-tactile bracelet. Right: One of our developers tests out the bracelets, which are worn on both arms.

Potential Applications
The VHP platform enables rapid experimentation and prototyping that can be used to develop techniques for a variety of applications. For example:

• Rich haptics on small devices: Expanding the number of actuators on mobile phones, which typically only have one or two, could be useful to provide additional tactile information. This is especially useful as fingers are sensitive to vibrations. We demonstrated a prototype mobile phone case with eight vibrotactile actuators. This could be used to provide rich notifications and enhance effects in a mobile game or when watching a video.
• Lab psychophysical experiments: Because VHP can be easily set up to send and receive haptic signals in real time, e.g., from a Jupyter notebook, it could be used to perform real-time haptic experiments.
• Notifications and alerts: The wearable VHP could be used to provide haptic notifications from other devices, e.g., alerting if someone is at the door, and could even communicate distinguishable alerts through use of multiple actuators.
• Sensory substitution: Besides the lipreading assistance example above, there are many other potential applications for accessibility using sensory substitution, such as visual-to-tactile sensing or even sensing magnetic fields.
• Loading sensing: The ability to sense from the haptic actuator current load is unique to our platform, and enables a variety of features, such as pressure sensing or automatically adjusting actuator output.

Integrating eight voice coils into a phone case. We used loading sensing to understand which voice coils are being touched.

What’s next?
We hope that others can utilize the platform to build a diverse set of applications. If you are interested and have ideas about using our platform or want to receive updates, please fill out this form. We hope that with this platform, we can help democratize the use of haptics and inspire a more widespread use of tactile devices.

Acknowledgments
This work was done by Artem Dementyev, Pascal Getreuer, Dimitri Kanevsky, Malcolm Slaney and Richard Lyon. We thank Alex Olwal, Thad Starner, Hong Tan, Charlotte Reed, Sarah Sterman for valuable feedback and discussion on the paper. Yuhui Zhao, Dmitrii Votintcev, Chet Gnegy, Whitney Bai and Sagar Savla for feedback on the design and engineering.

Posted by Dave Epstein, Student Researcher and Chen Sun, Staff Research Scientist, Google Research

Machine learning (ML) agents are increasingly deployed in the real world to make decisions and assist people in their daily lives. Making reasonable predictions about the future at varying timescales is one of the most important capabilities for such agents because it enables them to predict changes in the world around them, including other agents’ behaviors, and plan how to act next. Importantly, successful future prediction requires both capturing meaningful transitions in the environment (e.g., dough transforming into bread) and adapting to how transitions unfold over time in order to make decisions.

Previous work in future prediction from visual observations has largely been constrained by the format of its output (e.g., pixels that represent an image) or a manually-defined set of human activities (e.g., predicting if someone will keep walking, sit down, or jump). These are either too detailed and hard to predict or lack important information about the richness of the real world. For example, predicting “person jumping” does not capture why they’re jumping, what they’re jumping onto, etc. Also, with very few exceptions, previous models were designed to make predictions at a fixed offset into the future, which is a limiting assumption because we rarely know when meaningful future states will happen.

For example, in a video about making ice cream (depicted below), the meaningful transition from “cream” to “ice cream” occurs over 35 seconds, so models predicting such transitions would need to look 35 seconds ahead. But this time interval varies a large amount across different activities and videos — meaningful transitions occur at any distance into the future. Learning to make such predictions at flexible intervals is hard because the desired ground truth may be relatively ambiguous. For example, the correct prediction could be the just-churned ice cream in the machine, or scoops of the ice cream in a bowl. In addition, collecting such annotations at scale (i.e., frame-by-frame for millions of videos) is infeasible. However, many existing instructional videos come with speech transcripts, which often offer concise, general descriptions throughout entire videos. This source of data can guide a model’s attention toward important parts of the video, obviating the need for manual labeling and allowing a flexible, data-driven definition of the future.

In “Learning Temporal Dynamics from Cycles in Narrated Video”, published at ICCV 2021, we propose an approach that is self-supervised, using a recent large unlabeled dataset of diverse human action. The resulting model operates at a high level of abstraction, can make predictions arbitrarily far into the future, and chooses how far into the future to predict based on context. Called Multi-Modal Cycle Consistency (MMCC), it leverages narrated instructional video to learn a strong predictive model of the future. We demonstrate how MMCC can be applied, without fine-tuning, to a variety of challenging tasks, and qualitatively examine its predictions. In the example below, MMCC predicts the future (d) from present frame (a), rather than less relevant potential futures (b) or (c).

This work uses cues from vision and language to predict high-level changes (such as cream becoming ice cream) in video (video from HowTo100M).

Viewing Videos as Graphs
The foundation of our method is to represent narrated videos as graphs. We view videos as a collection of nodes, where nodes are either video frames (sampled at 1 frame per second) or segments of narrated text (extracted with automatic speech recognition systems), encoded by neural networks. During training, MMCC constructs a graph from the nodes, using cross-modal edges to connect video frames and text segments that refer to the same state, and temporal edges to connect the present (e.g., strawberry-flavored cream) and the future (e.g., soft-serve ice cream). The temporal edges operate on both modalities equally — they can start from either a video frame, some text, or both, and can connect to a future (or past) state in either modality. MMCC achieves this by learning a latent representation shared by frames and text and then making predictions in this representation space.

Multi-modal Cycle Consistency
To learn the cross-modal and temporal edge functions without supervision, we apply the idea of cycle consistency. Here, cycle consistency refers to the construction of cycle graphs, in which the model constructs a series of edges from an initial node to other nodes and back again: Given a start node (e.g., a sample video frame), the model is expected to find its cross-modal counterpart (i.e., text describing the frame) and combine them as the present state. To do this, at the start of training, the model assumes that frames and text with the same timestamps are counterparts, but then relaxes this assumption later. The model then predicts a future state, and the node most similar to this prediction is selected. Finally, the model attempts to invert the above steps by predicting the present state backward from the future node, and thus connecting the future node back with the start node.

The discrepancy between the model’s prediction of the present from the future and the actual present is the cycle-consistency loss. Intuitively, this training objective requires the predicted future to contain enough information about its past to be invertible, leading to predictions that correspond to meaningful changes to the same entities (e.g., tomato becoming marinara sauce, or flour and eggs in a bowl becoming dough). Moreover, the inclusion of cross-modal edges ensures future predictions are meaningful in either modality.

To learn the temporal and cross-modal edge functions end-to-end, we use the soft attention technique, which first outputs how likely each node is to be the target node of the edge, and then “picks” a node by taking the weighted average among all possible candidates. Importantly, this cyclic graph constraint makes few assumptions for the kind of temporal edges the model should learn, as long as they end up forming a consistent cycle. This enables the emergence of long-term temporal dynamics critical for future prediction without requiring manual labels of meaningful changes.

An example of the training objective: A cycle graph is expected to be constructed between the chicken with soy sauce and the chicken in chili oil because they are two adjacent steps in the chicken’s preparation (video from HowTo100M).

Discovering Cycles in Real-World Video
MMCC is trained without any explicit ground truth, using only long video sequences and randomly sampled starting conditions (a frame or text excerpt) and asking the model to find temporal cycles. After training, MMCC can identify meaningful cycles that capture complex changes in video.

Given frames as input (left), MMCC selects relevant text from video narrations and uses both modalities to predict a future frame (middle). It then finds text relevant to this future and uses it to predict the past (right). Using its knowledge of how objects and scenes change over time, MMCC “closes the cycle” and ends up where it started (videos from HowTo100M).

The model can also start from narrated text rather than frames and still find relevant transitions (videos from HowTo100M).

Zero-Shot Applications
For MMCC to identify meaningful transitions over time in an entire video, we define a “likely transition score” for each pair (A, B) of frames in a video, according to the model’s predictions — the closer B is to our model’s prediction of the future of A, the higher the score assigned. We then rank all pairs according to this score and show the highest-scoring pairs of present and future frames detected in previously unseen videos (examples below).

The highest-scoring pairs from eight random videos, which showcase the versatility of the model across a wide range of tasks (videos from HowTo100M).

We can use this same approach to temporally sort an unordered collection of video frames without any fine-tuning by finding an ordering that maximizes the overall confidence scores between all adjacent frames in the sorted sequence.

Left: Shuffled frames from three videos. Right: MMCC unshuffles the frames. The true order is shown under each frame. Even when MMCC does not predict the ground truth, its predictions often appear reasonable, and so, it can present an alternate ordering (videos from HowTo100M).

Evaluating Future Prediction
We evaluate the model’s ability to anticipate action, potentially minutes in advance, using the top-k recall metric, which here measures a model’s ability to retrieve the correct future (higher is better). On CrossTask, a dataset of instruction videos with labels describing key steps, MMCC outperforms the previous self-supervised state-of-the-art models in inferring possible future actions.

	Recall
Model	Top-1	Top-5	Top-10
Cross-modal	2.9	14.2	24.3
Repr. Ant.	3.0	13.3	26.0
MemDPC	2.9	15.8	27.4
TAP	4.5	17.1	27.9
MMCC	5.4	19.9	33.8

Conclusions
We have introduced a self-supervised method to learn temporal dynamics by cycling through narrated instructional videos. Despite the simplicity of the model’s architecture, it can discover meaningful long-term transitions in vision and language, and can be applied without further training to challenging downstream tasks, such as anticipating far-away action and ordering collections of images. An interesting future direction is transferring the model to agents so they can use it to conduct long-term planning.

Acknowledgements
The core team includes Dave Epstein, Jiajun Wu, Cordelia Schmid, and Chen Sun. We thank Alexei Efros, Mia Chiquier, and Shiry Ginosar for their feedback, and Allan Jabri for inspiration in figure design. Dave would like to thank Dídac Surís and Carl Vondrick for insightful early discussions on cycling through time in video.

Posted by Ye Jia and Michelle Tadmor Ramanovich, Software Engineers, Google Research

Speech-to-speech translation (S2ST) is key to breaking down language barriers between people all over the world. Automatic S2ST systems are typically composed of a cascade of speech recognition, machine translation, and speech synthesis subsystems. However, such cascade systems may suffer from longer latency, loss of information (especially paralinguistic and non-linguistic information), and compounding errors between subsystems.

In 2019, we introduced Translatotron, the first ever model that was able to directly translate speech between two languages. This direct S2ST model was able to be efficiently trained end-to-end and also had the unique capability of retaining the source speaker’s voice (which is non-linguistic information) in the translated speech. However, despite its ability to produce natural sounding translated speech in high fidelity, it still underperformed compared to a strong baseline cascade S2ST system (e.g., composed of a direct speech-to-text translation model [1, 2] followed by a Tacotron 2 TTS model).

In “Translatotron 2: Robust direct speech-to-speech translation”, we describe an improved version of Translatotron that significantly improves performance while also applying a new method for transferring the source speakers’ voices to the translated speech. The revised approach to voice transference is successful even when the input speech contains multiple speakers speaking in turns while also reducing the potential for misuse and better aligning with our AI Principles. Experiments on three different corpora consistently showed that Translatotron 2 outperforms the original Translatotron by a large margin on translation quality, speech naturalness, and speech robustness.

Translatotron 2
Translatotron 2 is composed of four major components: a speech encoder, a target phoneme decoder, a target speech synthesizer, and an attention module that connects them together. The combination of the encoder, the attention module, and the decoder is similar to a typical direct speech-to-text translation (ST) model. The synthesizer is conditioned on the output from both the decoder and the attention.

Model architecture of Translatotron 2 (for translating Spanish speech into English speech).

There are three novel changes between Translatotron and Translatotron 2 that are key factors in improving the performance:

1. While the output from the target phoneme decoder is used only as an auxiliary loss in the original Translatotron, it is one of the inputs to the spectrogram synthesizer in Translatotron 2. This strong conditioning makes Translatotron 2 easier to train and yields better performance.
2. The spectrogram synthesizer in the original Translatotron is attention-based, similar to the Tacotron 2 TTS model, and as a consequence, it also suffers from the robustness issues exhibited by Tacotron 2. In contrast, the spectrogram synthesizer employed in Translatotron 2 is duration-based, similar to that used by Non-Attentive Tacotron, which drastically improves the robustness of the synthesized speech.
3. Both Translatotron and Translatotron 2 use an attention-based connection to the encoded source speech. However, in Translatotron 2, this attention is driven by the phoneme decoder instead of the spectrogram synthesizer. This ensures the acoustic information that the spectrogram synthesizer sees is aligned with the translated content that it’s synthesizing, which helps retain each speaker’s voice across speaker turns.

More Powerful and Responsible Voice Retention
The original Translatotron was able to retain the source speaker’s voice in the translated speech, by conditioning its decoder on a speaker embedding generated from a separately trained speaker encoder. However, this approach also enabled it to generate the translated speech in a different speaker’s voice if a clip of the target speaker’s recording were used as the reference audio to the speaker encoder, or if the embedding of the target speaker were directly available. While this capability was powerful, it had the potential to be misused to spoof audio with arbitrary content, which posed a concern for production deployment.

To address this, we designed Translatotron 2 to use only a single speech encoder, which is responsible for both linguistic understanding and voice capture. In this way, the trained models cannot be directed to reproduce non-source voices. This approach can also be applied to the original Translatotron.

To retain speakers’ voices across translation, researchers generally prefer to train S2ST models on parallel utterances with the same speaker’s voice on both sides. Such a dataset with human recordings on both sides is extremely difficult to collect, because it requires a large number of fluent bilingual speakers. To avoid this difficulty, we use a modified version of PnG NAT, a TTS model that is capable of cross-lingual voice transferring to synthesize such training targets. Our modified PnG NAT model incorporates a separately trained speaker encoder in the same way as in our previous TTS work — the same strategy used for the original Translatotron — so that it is capable of zero-shot voice transference.

Following are examples of direct speech-to-speech translation from Translatotron 2 in which the source speaker’s voice is retained:

Input (Spanish):

TTS-synthesized reference (English):

Translatotron 2 prediction (English):

Translatotron prediction (English):

To enable S2ST models to retain each speaker’s voice in the translated speech when the input speech contains multiple speakers speaking in turns, we propose a simple concatenation-based data augmentation technique, called ConcatAug. This method augments the training data on the fly by randomly sampling pairs of training examples and concatenating the source speech, the target speech, and the target phoneme sequences into new training examples. The resulting samples contain two speakers’ voices in both the source and the target speech, which enables the model to learn on examples with speaker turns. Following are audio samples from Translatotron 2 with speaker turns:

Input (Spanish):

TTS-synthesized reference (English):

Translatotron 2 (with ConcatAug) prediction (English):

Translatotron 2 (without ConcatAug) prediction (English):

More audio samples are available here.

Performance
Translatotron 2 outperforms the original Translatotron by large margins in every aspect we measured: higher translation quality (measured by BLEU, where higher is better), speech naturalness (measured by MOS, higher is better), and speech robustness (measured by UDR, lower is better). It particularly excelled on the more difficult Fisher corpus. The performance of Translatotron 2 on translation quality and speech quality approaches that of a strong baseline cascade system, and is better than the cascade baseline on speech robustness.

Translation quality (measured by BLEU, where higher is better) evaluated on two Spanish-English corpora.

Speech naturalness (measured by MOS, where higher is better) evaluated on two Spanish-English corpora.

Speech robustness (measured by UDR, where lower is better) evaluated on two Spanish-English corpora.

Multilingual Speech-to-Speech Translation
Besides Spanish-to-English S2ST, we also evaluated the performance of Translatotron 2 on a multilingual set-up in which the model took speech input from four different languages and translated them into English. The language of the input speech was not provided, which forced the model to detect the language by itself.

Source Language	fr	de	es	ca
Translatotron 2	27.0	18.8	27.7	22.5
Translatotron	18.9	10.8	18.8	13.9
ST (Wang et al. 2020)	27.0	18.9	28.0	23.9
Training Target	82.1	86.0	85.1	89.3

Performance of multilingual X=>En S2ST on the CoVoST 2 corpus.

On this task, Translatotron 2 again outperformed the original Translatotron by a large margin. Although the results are not directly comparable between S2ST and ST, the close numbers suggest that the translation quality from Translatotron 2 is comparable to a baseline speech-to-text translation model, These results indicate that Translatotron 2 is also highly effective on multilingual S2ST.

Acknowledgments
The direct contributors to this work include Ye Jia, Michelle Tadmor Ramanovich, Tal Remez, Roi Pomerantz. We also thank Chung-Cheng Chiu, Quan Wang, Heiga Zen, Ron J. Weiss, Wolfgang Macherey, Yu Zhang, Yonghui Wu, Hadar Shemtov, Ruoming Pang, Nadav Bar, Hen Fitoussi, Benny Schlesinger, Michael Hassid for helpful discussions and support.

Posted by Jing Yu Koh, Research Engineer and Peter Anderson, Senior Research Scientist, Google Research

When a person navigates around an unfamiliar building, they take advantage of many visual, spatial and semantic cues to help them efficiently reach their goal. For example, even in an unfamiliar house, if they see a dining area, they can make intelligent predictions about the likely location of the kitchen and lounge areas, and therefore the expected location of common household objects. For robotic agents, taking advantage of semantic cues and statistical regularities in novel buildings is challenging. A typical approach is to implicitly learn what these cues are, and how to use them for navigation tasks, in an end-to-end manner via model-free reinforcement learning. However, navigation cues learned in this way are expensive to learn, hard to inspect, and difficult to re-use in another agent without learning again from scratch.

People navigating in unfamiliar buildings can take advantage of visual, spatial and semantic cues to predict what’s around a corner. A computational model with this capability is a visual world model.

An appealing alternative for robotic navigation and planning agents is to use a world model to encapsulate rich and meaningful information about their surroundings, which enables an agent to make specific predictions about actionable outcomes within their environment. Such models have seen widespread interest in robotics, simulation, and reinforcement learning with impressive results, including finding the first known solution for a simulated 2D car racing task, and achieving human-level performance in Atari games. However, game environments are still relatively simple compared to the complexity and diversity of real-world environments.

In “Pathdreamer: A World Model for Indoor Navigation”, published at ICCV 2021, we present a world model that generates high-resolution 360º visual observations of areas of a building unseen by an agent, using only limited seed observations and a proposed navigation trajectory. As illustrated in the video below, the Pathdreamer model can synthesize an immersive scene from a single viewpoint, predicting what an agent might see if it moved to a new viewpoint or even a completely unseen area, such as around a corner. Beyond potential applications in video editing and bringing photos to life, solving this task promises to codify knowledge about human environments to benefit robotic agents navigating in the real world. For example, a robot tasked with finding a particular room or object in an unfamiliar building could perform simulations using the world model to identify likely locations before physically searching anywhere. World models such as Pathdreamer can also be used to increase the amount of training data for agents, by training agents in the model.

Provided with just a single observation (RGB, depth, and segmentation) and a proposed navigation trajectory as input, Pathdreamer synthesizes high resolution 360º observations up to 6-7 meters away from the original location, including around corners. For more results, please refer to the full video.

How Does Pathdreamer Work?
Pathdreamer takes as input a sequence of one or more previous observations, and generates predictions for a trajectory of future locations, which may be provided up front or iteratively by the agent interacting with the returned observations. Both inputs and predictions consist of RGB, semantic segmentation, and depth images. Internally, Pathdreamer uses a 3D point cloud to represent surfaces in the environment. Points in the cloud are labelled with both their RGB color value and their semantic segmentation class, such as wall, chair or table.

To predict visual observations in a new location, the point cloud is first re-projected into 2D at the new location to provide ‘guidance’ images, from which Pathdreamer generates realistic high-resolution RGB, semantic segmentation and depth. As the model ‘moves’, new observations (either real or predicted) are accumulated in the point cloud. One advantage of using a point cloud for memory is temporal consistency — revisited regions are rendered in a consistent manner to previous observations.

Internally, Pathdreamer represents surfaces in the environment via a 3D point cloud containing both semantic labels (top) and RGB color values (bottom). To generate a new observation, Pathdreamer ‘moves’ through the point cloud to the new location and uses the re-projected point cloud image for guidance.

To convert guidance images into plausible, realistic outputs Pathdreamer operates in two stages: the first stage, the structure generator, creates segmentation and depth images, and the second stage, the image generator, renders these into RGB outputs. Conceptually, the first stage provides a plausible high-level semantic representation of the scene, and the second stage renders this into a realistic color image. Both stages are based on convolutional neural networks.

Pathdreamer operates in two stages: the first stage, the structure generator, creates segmentation and depth images, and the second stage, the image generator, renders these into RGB outputs. The structure generator is conditioned on a noise variable to enable the model to synthesize diverse scenes in areas of high uncertainty.

Diverse Generation Results
In regions of high uncertainty, such as an area predicted to be around a corner or in an unseen room, many different scenes are possible. Incorporating ideas from stochastic video generation, the structure generator in Pathdreamer is conditioned on a noise variable, which represents the stochastic information about the next location that is not captured in the guidance images. By sampling multiple noise variables, Pathdreamer can synthesize diverse scenes, allowing an agent to sample multiple plausible outcomes for a given trajectory. These diverse outputs are reflected not only in the first stage outputs (semantic segmentation and depth images), but in the generated RGB images as well.

Pathdreamer is capable of generating multiple diverse and plausible images for regions of high uncertainty. Guidance images on the leftmost column represent pixels that were previously seen by the agent. Black pixels represent regions that were previously unseen, for which Pathdreamer renders diverse outputs by sampling multiple random noise vectors. In practice, the generated output can be informed by new observations as the agent navigates the environment.

<!–

Pathdreamer is capable of generating multiple diverse and plausible images for regions of high uncertainty. Guidance images on the leftmost column represent pixels that were previously seen by the agent. Black pixels represent regions that were previously unseen, for which Pathdreamer renders diverse outputs by sampling multiple random noise vectors. In practice, the generated output can be informed by new observations as the agent navigates the environment.

–>

Pathdreamer is trained with images and 3D environment reconstructions from Matterport3D, and is capable of synthesizing realistic images as well as continuous video sequences. Because the output imagery is high-resolution and 360º, it can be readily converted for use by existing navigation agents for any camera field of view. For more details and to try out Pathdreamer yourself, we recommend taking a look at our open source code.

Application to Visual Navigation Tasks
As a visual world model, Pathdreamer shows strong potential to improve performance on downstream tasks. To demonstrate this, we apply Pathdreamer to the task of Vision-and-Language Navigation (VLN), in which an embodied agent must follow a natural language instruction to navigate to a location in a realistic 3D environment. Using the Room-to-Room (R2R) dataset, we conduct an experiment in which an instruction-following agent plans ahead by simulating many possible navigable trajectory through the environment, ranking each against the navigation instructions, and choosing the best ranked trajectory to execute. Three settings are considered. In the Ground-Truth setting, the agent plans by interacting with the actual environment, i.e. by moving. In the Baseline setting, the agent plans ahead without moving by interacting with a navigation graph that encodes the navigable routes within the building, but does not provide any visual observations. In the Pathdreamer setting, the agent plans ahead without moving by interacting with the navigation graph and also receives corresponding visual observations generated by Pathdreamer.

When planning ahead for three steps (approximately 6m), in the Pathdreamer setting the VLN agent achieves a navigation success rate of 50.4%, significantly higher than the 40.6% success rate in the Baseline setting without Pathdreamer. This suggests that Pathdreamer encodes useful and accessible visual, spatial and semantic knowledge about real-world indoor environments. As an upper bound illustrating the performance of a perfect world model, under the Ground-Truth setting (planning by moving) the agent’s success rate is 59%, although we note that this setting requires the agent to expend significant time and resources to physically explore many trajectories, which would likely be prohibitively costly in a real-world setting.

We evaluate several planning settings for an instruction-following agent using the Room-to-Room (R2R) dataset. Planning ahead using a navigation graph with corresponding visual observations synthesized by Pathdreamer (Pathdreamer setting) is more effective than planning ahead using the navigation graph alone (Baseline setting), capturing around half the benefit of planning ahead using a world model that perfectly matches reality (Ground-Truth setting).

Conclusions and Future Work
These results showcase the promise of using world models such as Pathdreamer for complicated embodied navigation tasks. We hope that Pathdreamer will help unlock model-based approaches to challenging embodied navigation tasks such as navigating to specified objects and VLN.

Applying Pathdreamer to other embodied navigation tasks such as Object-Nav, continuous VLN, and street-level navigation are natural directions for future work. We also envision further research on improved architecture and modeling directions for the Pathdreamer model, as well as testing it on more diverse datasets, including but not limited to outdoor environments. To explore Pathdreamer in more detail, please visit our GitHub repository.

Acknowledgements
This project is a collaboration with Jason Baldridge, Honglak Lee, and Yinfei Yang. We thank Austin Waters, Noah Snavely, Suhani Vora, Harsh Agrawal, David Ha, and others who provided feedback throughout the project. We are also grateful for general support from Google Research teams. Finally, we thank Tom Small for creating the animation in the third figure.

Posted by Krishna Srinivasan, Software Engineer and Karthik Raman, Research Scientist, Google Research

Multimodal visio-linguistic models rely on rich datasets in order to model the relationship between images and text. Traditionally, these datasets have been created by either manually captioning images, or crawling the web and extracting the alt-text as the caption. While the former approach tends to result in higher quality data, the intensive manual annotation process limits the amount of data that can be created. On the other hand, the automated extraction approach can lead to bigger datasets, but these require either heuristics and careful filtering to ensure data quality or scaling-up models to achieve strong performance. An additional shortcoming of existing datasets is the dearth of coverage in non-English languages. This naturally led us to ask: Can one overcome these limitations and create a high-quality, large-sized, multilingual dataset with a variety of content?

Today we introduce the Wikipedia-Based Image Text (WIT) Dataset, a large multimodal dataset, created by extracting multiple different text selections associated with an image from Wikipedia articles and Wikimedia image links. This was accompanied by rigorous filtering to only retain high quality image-text sets. As detailed in “WIT: Wikipedia-based Image Text Dataset for Multimodal Multilingual Machine Learning”, presented at SIGIR ‘21, this resulted in a curated set of 37.5 million entity-rich image-text examples with 11.5 million unique images across 108 languages. The WIT dataset is available for download and use under the Creative Commons license. We are also excited to announce that we are hosting a competition with the WIT dataset in Kaggle in collaboration with Wikimedia Research and other external collaborators.

Dataset	Images	Text	Contextual Text	Languages
Flickr30K	32K	158K	–	< 8
SBU Captions	1M	1M	–	1
MS-COCO	330K	1.5M	–	< 4; 7 (test only)
CC-3M CC-12M	3.3M 12M	3.3M 12M	– –	1 1
WIT	11.5M	37.5M	~119M	108

WIT’s increased language coverage and larger size relative to previous datasets.

The unique advantages of the WIT dataset are:

1. Size: WIT is the largest multimodal dataset of image-text examples that is publicly available.
2. Multilingual: With 108 languages, WIT has 10x or more languages than any other dataset.
3. Contextual information: Unlike typical multimodal datasets, which have only one caption per image, WIT includes many page-level and section-level contextual information.
4. Real world entities: Wikipedia, being a broad knowledge-base, is rich with real world entities that are represented in WIT.
5. Challenging test set: In our recent work accepted at EMNLP, all state-of-the-art models demonstrated significantly lower performance on WIT vs. traditional evaluation sets (e.g., ~30 point drop in recall).

Generating the Dataset
The main goal of WIT was to create a large dataset without sacrificing on quality or coverage of concepts. Thus, we started by leveraging the largest online encyclopedia available today: Wikipedia.

For an example of the depth of information available, consider the Wikipedia page for Half Dome (Yosemite National Park, CA). As shown below, the article has numerous interesting text captions and relevant contextual information for the image, such as the page title, main page description, and other contextual information and metadata.

Example wikipedia page with various image-associated text selections and contexts we can extract. From the Wikipedia page for Half Dome : Photo by DAVID ILIFF. License: CC BY-SA 3.0.

Example of the Wikipedia page for this specific image of Half Dome. From the Wikipedia page for Half Dome : Photo by DAVID ILIFF. License: CC BY-SA 3.0.

We started by selecting Wikipedia pages that have images, then extracted various image-text associations and surrounding contexts. To further refine the data, we performed a rigorous filtering process to ensure data quality. This included text-based filtering to ensure caption availability, length and quality (e.g., by removing generic default filler text); image-based filtering to ensure each image is a certain size with permissible licensing; and finally, image-and-text-entity–based filtering to ensure suitability for research (e.g., excluding those classified as hate speech). We further randomly sampled image-caption sets for evaluation by human editors, who overwhelmingly agreed that 98% of the samples had good image-caption alignment.

Highly Multilingual
With data in 108 languages, WIT is the first large-scale, multilingual, multimodal dataset.

# of Image-Text Sets	Unique Languages	# of Images	Unique Languages
> 1M	9	> 1M	6
500K – 1M	10	500K – 1M	12
100K – 500K	36	100K – 500K	35
50K – 100K	15	50K – 100K	17
14K – 50K	38	13K – 50K	38

WIT: coverage statistics across languages.

Example of an image that is present in more than a dozen Wikipedia pages across >12 languages. From the Wikipedia page for Wolfgang Amadeus Mozart.

The First Contextual Image-Text Dataset
Most multimodal datasets only offer a single text caption (or multiple versions of a similar caption) for the given image. WIT is the first dataset to provide contextual information, which can help researchers model the effect of context on image captions as well as the choice of images.

WIT dataset example showing image-text data and additional contextual information.

In particular, key textual fields of WIT that may be useful for research include:

• Text captions: WIT offers three different kinds of image captions. This includes the (potentially context influenced) “Reference description”, the (likely context independent) “Attribution description” and “Alt-text description”.
• Contextual information: This includes the page title, page description, URL and local context about the Wikipedia section including the section title and text.

WIT has broad coverage across these different fields, as shown below.

Image-Text Fields of WIT	Train	Val	Test	Total / Unique
Rows / Tuples	37.1M	261.8K	210.7K	37.6M
Unique Images	11.4M	58K	57K	11.5M
Reference Descriptions	16.9M	150K	104K	17.2M / 16.7M
Attribution Descriptions	34.8M	193K	200K	35.2M / 10.9M
Alt-Text	5.3M	29K	29K	5.4M / 5.3M
Context Texts	–	–	–	119.8M

Key fields of WIT include both text captions and contextual information.

A High-Quality Training Set and a Challenging Evaluation Benchmark
The broad coverage of diverse concepts in Wikipedia means that the WIT evaluation sets serve as a challenging benchmark, even for state-of-the-art models. We found that for image-text retrieval, the mean recall scores for traditional datasets were in the 80s, whereas for the WIT test set, it was in the 40s for well-resourced languages and in the 30s for the under-resourced languages. We hope this in turn can help researchers to build stronger, more robust models.

WIT Dataset and Competition with Wikimedia and Kaggle
Additionally, we are happy to announce that we are partnering with Wikimedia Research and a few external collaborators to organize a competition with the WIT test set. We are hosting this competition in Kaggle. The competition is an image-text retrieval task. Given a set of images and text captions, the task is to retrieve the appropriate caption(s) for each image.

To enable research in this area, Wikipedia has kindly made available images at 300-pixel resolution and a Resnet-50–based image embeddings for most of the training and the test dataset. Kaggle will be hosting all this image data in addition to the WIT dataset itself and will provide colab notebooks. Further, the competitors will have access to a discussion forum in Kaggle in order to share code and collaborate. This enables anyone interested in multimodality to get started and run experiments easily. We are excited and looking forward to what will result from the WIT dataset and the Wikipedia images in the Kaggle platform.

Conclusion
We believe that the WIT dataset will aid researchers in building better multimodal multilingual models and in identifying better learning and representation techniques, ultimately leading to improved Machine Learning models in real-world tasks over visio-linguistic data. For any questions, please contact wit-dataset@google.com. We would love to hear about how you are using the WIT dataset.

Acknowledgements
We would like to thank our co-authors in Google Research: Jiecao Chen, Michael Bendersky and Marc Najork. We thank Beer Changpinyo, Corinna Cortes, Joshua Gang, Chao Jia, Ashwin Kakarla, Mike Lee, Zhen Li, Piyush Sharma, Radu Soricut, Ashish Vaswani, Yinfei Yang, and our reviewers for their insightful feedback and comments.

We thank Miriam Redi and Leila Zia from Wikimedia Research for collaborating with us on the competition and providing image pixels and image embedding data. We thank Addison Howard and Walter Reade for helping us host this competition in Kaggle. We also thank Diane Larlus (Naver Labs Europe (NLE)), Yannis Kalantidis (NLE), Stéphane Clinchant (NLE), Tiziano Piccardi Ph.D. student at EPFL, Lucie-Aimée Kaffee PhD student at University of Southampton and Yacine Jernite (Hugging Face) for their valuable contribution towards the competition.

Posted by Mingxing Tan and Zihang Dai, Research Scientists, Google Research

As neural network models and training data size grow, training efficiency is becoming an important focus for deep learning. For example, GPT-3 demonstrates remarkable capability in few-shot learning, but it requires weeks of training with thousands of GPUs, making it difficult to retrain or improve. What if, instead, one could design neural networks that were smaller and faster, yet still more accurate?

In this post, we introduce two families of models for image recognition that leverage neural architecture search, and a principled design methodology based on model capacity and generalization. The first is EfficientNetV2 (accepted at ICML 2021), which consists of convolutional neural networks that aim for fast training speed for relatively small-scale datasets, such as ImageNet1k (with 1.28 million images). The second family is CoAtNet, which are hybrid models that combine convolution and self-attention, with the goal of achieving higher accuracy on large-scale datasets, such as ImageNet21 (with 13 million images) and JFT (with billions of images). Compared to previous results, our models are 4-10x faster while achieving new state-of-the-art 90.88% top-1 accuracy on the well-established ImageNet dataset. We are also releasing the source code and pretrained models on the Google AutoML github.

EfficientNetV2: Smaller Models and Faster Training
EfficientNetV2 is based upon the previous EfficientNet architecture. To improve upon the original, we systematically studied the training speed bottlenecks on modern TPUs/GPUs and found: (1) training with very large image sizes results in higher memory usage and thus is often slower on TPUs/GPUs; (2) the widely used depthwise convolutions are inefficient on TPUs/GPUs, because they exhibit low hardware utilization; and (3) the commonly used uniform compound scaling approach, which scales up every stage of convolutional networks equally, is sub-optimal. To address these issues, we propose both a training-aware neural architecture search (NAS), in which the training speed is included in the optimization goal, and a scaling method that scales different stages in a non-uniform manner.

The training-aware NAS is based on the previous platform-aware NAS, but unlike the original approach, which mostly focuses on inference speed, here we jointly optimize model accuracy, model size, and training speed. We also extend the original search space to include more accelerator-friendly operations, such as FusedMBConv, and simplify the search space by removing unnecessary operations, such as average pooling and max pooling, which are never selected by NAS. The resulting EfficientNetV2 networks achieve improved accuracy over all previous models, while being much faster and up to 6.8x smaller.

To further speed up the training process, we also propose an enhanced method of progressive learning, which gradually changes image size and regularization magnitude during training. Progressive training has been used in image classification, GANs, and language models. This approach focuses on image classification, but unlike previous approaches that often trade accuracy for improved training speed, can slightly improve the accuracy while also significantly reducing training time. The key idea in our improved approach is to adaptively change regularization strength, such as dropout ratio or data augmentation magnitude, according to the image size. For the same network, small image size leads to lower network capacity and thus requires weak regularization; vice versa, a large image size requires stronger regularization to combat overfitting.

Progressive learning for EfficientNetV2. Here we mainly focus on three types of regularizations: data augmentation, mixup, and dropout.

We evaluate the EfficientNetV2 models on ImageNet and a few transfer learning datasets, such as CIFAR-10/100, Flowers, and Cars. On ImageNet, EfficientNetV2 significantly outperforms previous models with about 5–11x faster training speed and up to 6.8x smaller model size, without any drop in accuracy.

EfficientNetV2 achieves much better training efficiency than prior models for ImageNet classification.

CoAtNet: Fast and Accurate Models for Large-Scale Image Recognition
While EfficientNetV2 is still a typical convolutional neural network, recent studies on Vision Transformer (ViT) have shown that attention-based transformer models could perform better than convolutional neural networks on large-scale datasets like JFT-300M. Inspired by this observation, we further expand our study beyond convolutional neural networks with the aim of finding faster and more accurate vision models.

In “CoAtNet: Marrying Convolution and Attention for All Data Sizes”, we systematically study how to combine convolution and self-attention to develop fast and accurate neural networks for large-scale image recognition. Our work is based on an observation that convolution often has better generalization (i.e., the performance gap between training and evaluation) due to its inductive bias, while self-attention tends to have greater capacity (i.e., the ability to fit large-scale training data) thanks to its global receptive field. By combining convolution and self-attention, our hybrid models can achieve both better generalization and greater capacity.

Comparison between convolution, self-attention, and hybrid models. Convolutional models converge faster, ViTs have better capacity, while the hybrid models achieve both faster convergence and better accuracy.

We observe two key insights from our study: (1) depthwise convolution and self-attention can be naturally unified via simple relative attention, and (2) vertically stacking convolution layers and attention layers in a way that considers their capacity and computation required in each stage (resolution) is surprisingly effective in improving generalization, capacity and efficiency. Based on these insights, we have developed a family of hybrid models with both convolution and attention, named CoAtNets (pronounced “coat” nets). The following figure shows the overall CoAtNet network architecture:

Overall CoAtNet architecture. Given an input image with size HxW, we first apply convolutions in the first stem stage (S0) and reduce the size to H/2 x W/2. The size continues to reduce with each stage. Ln refers to the number of layers. Then, the early two stages (S1 and S2) mainly adopt MBConv building blocks consisting of depthwise convolution. The later two stages (S3 and S4) mainly adopt Transformer blocks with relative self-attention. Unlike the previous Transformer blocks in ViT, here we use pooling between stages, similar to Funnel Transformer. Finally, we apply a classification head to generate class prediction.

CoAtNet models consistently outperform ViT models and its variants across a number of datasets, such as ImageNet1K, ImageNet21K, and JFT. When compared to convolutional networks, CoAtNet exhibits comparable performance on a small-scale dataset (ImageNet1K) and achieves substantial gains as the data size increases (e.g. on ImageNet21K and JFT).

Comparison between CoAtNet and previous models after pre-training on the medium sized ImageNet21K dataset. Under the same model size, CoAtNet consistently outperforms both ViT and convolutional models. Noticeably, with only ImageNet21K, CoAtNet is able to match the performance of ViT-H pre-trained on JFT.

We also evaluated CoAtNets on the large-scale JFT dataset. To reach a similar accuracy target, CoAtNet trains about 4x faster than previous ViT models and more importantly, achieves a new state-of-the-art top-1 accuracy on ImageNet of 90.88%.

Comparison between CoAtNets and previous ViTs. ImageNet top-1 accuracy after pre-training on JFT dataset under different training budget. The four best models are trained on JFT-3B with about 3 billion images.

Conclusion and Future Work
In this post, we introduce two families of neural networks, named EfficientNetV2 and CoAtNet, which achieve state-of-the-art performance on image recognition. All EfficientNetV2 models are open sourced and the pretrained models are also available on the TFhub. CoAtNet models will also be open-sourced soon. We hope these new neural networks can benefit the research community and the industry. In the future we plan to further optimize these models and apply them to new tasks, such as zero-shot learning and self-supervised learning, which often require fast models with high capacity.

Acknowledgements
Special thanks to our co-authors Hanxiao Liu and Quoc Le. We also thank the Google Research, Brain Team and the open source contributors.

Posted by Vighnesh Birodkar, Research Software Engineer and Jonathan Huang, Research Scientist, Google Research

Instance segmentation is the task of grouping pixels in an image into instances of individual things, and identifying those things with a class label (countable objects such as people, animals, cars, etc., and assigning unique identifiers to each, e.g., car_1 and car_2). As a core computer vision task, it is critical to many downstream applications, such as self-driving cars, robotics, medical imaging, and photo editing. In recent years, deep learning has made significant strides in solving the instance segmentation problem with architectures like Mask R-CNN. However, these methods rely on collecting a large labeled instance segmentation dataset. But unlike bounding box labels, which can be collected in 7 seconds per instance with methods like Extreme clicking, collecting instance segmentation labels (called “masks”) can take up to 80 seconds per instance, an effort that is costly and creates a high barrier to entry for this research. And a related task, pantopic segmentation, requires even more labeled data.

The partially supervised instance segmentation setting, where only a small set of classes are labeled with instance segmentation masks and the remaining (majority of) classes are labeled only with bounding boxes, is an approach that has the potential to reduce the dependence on manually-created mask labels, thereby significantly lowering the barriers to developing an instance segmentation model. However this partially supervised approach also requires a stronger form of model generalization to handle novel classes not seen at training time—e.g., training with only animal masks and then tasking the model to produce accurate instance segmentations for buildings or plants. Further, naïve approaches, such as training a class-agnostic Mask R-CNN, while ignoring mask losses for any instances that don’t have mask labels, have not worked well. For example, on the typical “VOC/Non-VOC” benchmark, where one trains on masks for a subset of 20 classes in COCO (called “seen classes”) and is tested on the remaining 60 classes (called “unseen classes”), a typical Mask R-CNN with Resnet-50 backbone gets to only ~18% mask mAP (mean Average Precision, higher is better) on unseen classes, whereas when fully supervised it can achieve a much higher >34% mask mAP on the same set.

In “The surprising impact of mask-head architecture on novel class segmentation”, to be presented at ICCV 2021, we identify the main culprits for Mask R-CNN’s poor performance on novel classes and propose two easy-to-implement fixes (one training protocol fix, one mask-head architecture fix) that work in tandem to close the gap to fully supervised performance. We show that our approach applies generally to crop-then-segment models, i.e., a Mask R-CNN or Mask R-CNN-like architecture that computes a feature representation of the entire image and then subsequently passes per-instance crops to a second-stage mask prediction network—also called a mask-head network. Putting our findings together, we propose a Mask R-CNN–based model that improves over the current state-of-the-art by a significant 4.7% mask mAP without requiring more complex auxiliary loss functions, offline trained priors, or weight transfer functions proposed by previous work. We have also open sourced the code bases for two versions of the model, called Deep-MAC and Deep-MARC, and published a colab to interactively produce masks like the video demo below.

A demo of our model, DeepMAC, which learns to predict accurate masks, given user specified boxes, even on novel classes that were not seen at training time. Try it yourself in the colab. Image credits: Chris Briggs, Wikipedia and Europeana.

Impact of Cropping Methodology in Partially Supervised Settings
An important step of crop-then-segment models is cropping—Mask R-CNN is trained by cropping a feature map as well as the ground truth mask to a bounding box corresponding to each instance. These cropped features are passed to another neural network (called a mask-head network) that computes a final mask prediction, which is then compared against the ground truth crop in the mask loss function. There are two choices for cropping: (1) cropping directly to the ground truth bounding box of an instance, or (2) cropping to bounding boxes predicted by the model (called, proposals). At test time, cropping is always performed with proposals as ground truth boxes are not assumed to be available.

Cropping to ground truth boxes vs. cropping to proposals predicted by a model during training. Standard Mask R-CNN implementations use both types of crops, but we show that cropping exclusively to ground truth boxes yields significantly stronger performance on novel categories.

We consider a general family of Mask R-CNN–like architectures with one small, but critical difference from typical Mask R-CNN training setups: we crop using ground truth boxes (instead of proposal boxes) at training time.

Typical Mask R-CNN implementations pass both types of crops to the mask head. However, this choice has traditionally been considered an unimportant implementation detail, because it does not affect performance significantly in the fully supervised setting. In contrast, for partially supervised settings, we find that cropping methodology plays a significant role—while cropping exclusively to ground truth boxes during training doesn’t change the results significantly in the fully supervised setting, it has a surprising and dramatic positive impact in the partially supervised setting, performing significantly better on unseen classes.

Performance of Mask R-CNN on unseen classes when trained with either proposals and ground truth (the default) or with only ground truth boxes. Training mask heads with only ground truth boxes yields a significant boost to performance on unseen classes, upwards of 9% mAP. We report performance with the ResNet-101-FPN backbone.

Unlocking the Full Generalization Potential of the Mask Head
Even more surprisingly, the above approach unlocks a novel phenomenon—with cropping-to-ground truth enabled during training, the mask head of Mask R-CNN takes on a disproportionate role in the ability of the model to generalize to unseen classes. As an example, in the following figure, we compare models that all have cropping-to-ground-truth enabled, but different out-of-the-box mask-head architectures on a parking meter, cell phone, and pizza (classes unseen during training).

Mask predictions for unseen classes with four different mask-head architectures (from left to right: ResNet-4, ResNet-12, ResNet-20, Hourglass-20, where the number refers to the number of layers of the neural network). Despite never having seen masks from the ‘parking meter’, ‘pizza’ or ‘mobile phone’ class, the rightmost mask-head architecture can segment these classes correctly. From left to right, we show better mask-head architectures predicting better masks. Moreover, this difference is only apparent when evaluating on unseen classes — if we evaluate on seen classes, all four architectures exhibit similar performance.

Particularly notable is that these differences between mask-head architectures are not as obvious in the fully supervised setting. Incidentally, this may explain why previous works in instance segmentation have almost exclusively used shallow (i.e., low number of layers) mask heads, as there has been no benefit to the added complexity. Below we compare the mask mAP of three different mask-head architectures on seen versus unseen classes. All three models do equally well on the set of seen classes, but the deep hourglass mask heads stand out when applied to unseen classes. We find hourglass mask heads to be the best among the architectures we tried and we use hourglass mask heads with 50 or more layers to get the best results.

Performance of ResNet-4, Hourglass-10 and Hourglass-52 mask-head architectures on seen and unseen classes. There is a significant difference in performance on unseen classes, even though the performance on seen classes barely changes.

Finally, we show that our findings are general, holding for a variety of backbones (e.g., ResNet, SpineNet, Hourglass) and detector architectures including anchor-based and anchor-free detectors and even when there is no detector at all.

Putting It Together
To achieve the best result, we combined the above findings: We trained a Mask R-CNN model with cropping-to-ground-truth enabled and a deep Hourglass-52 mask head with a SpineNet backbone on high resolution images (1280×1280). We call this model Deep-MARC (Deep Mask heads Above R–CNN). Without using any offline training or other hand-crafted priors, Deep-MARC exceeds previous state-of-the-art models by > 4.5% (absolute) mask mAP. Demonstrating the general nature of this approach, we also see strong results with a CenterNet-based (as opposed to Mask R-CNN-based) model (called Deep-MAC), which also exceeds the previous state of the art.

Comparison of Deep-MAC and Deep-MARC to other partially supervised instance segmentation approaches like MaskX R-CNN, ShapeMask and CPMask.

Conclusion
We develop instance segmentation models that are able to generalize to classes that were not part of the training set. We highlight the role of two key ingredients that can be applied to any crop-then-segment model (such as Mask R-CNN): (1) cropping-to-ground truth boxes during training, and (2) strong mask-head architectures. While neither of these ingredients have a large impact on the classes for which masks are available during training, employing both leads to significant improvement on novel classes for which masks are not available during training. Moreover, these ingredients are sufficient for achieving state-of-the-art-performance on the partially-supervised COCO benchmark. Finally, our findings are general and may also have implications for related tasks, such as panoptic segmentation and pose estimation.

Acknowledgements
We thank our co-authors Zhichao Lu, Siyang Li, and Vivek Rathod. We thank David Ross and our anonymous ICCV reviewers for their comments which played a big part in improving this research.

Posted by Shan Yang, Software Engineer and Angjoo Kanazawa, Research Scientist, Google Research

Dancing is a universal language found in nearly all cultures, and is an outlet many people use to express themselves on contemporary media platforms today. The ability to dance by composing movement patterns that align to music beats is a fundamental aspect of human behavior. However, dancing is a form of art that requires practice. In fact, professional training is often required to equip a dancer with a rich repertoire of dance motions needed to create expressive choreography. While this process is difficult for people, it is even more challenging for a machine learning (ML) model, because the task requires the ability to generate a continuous motion with high kinematic complexity, while capturing the non-linear relationship between the movements and the accompanying music.

In “AI Choreographer: Music-Conditioned 3D Dance Generation with AIST++”, presented at ICCV 2021, we propose a full-attention cross-modal Transformer (FACT) model can mimic and understand dance motions, and can even enhance a person’s ability to choreograph dance. Together with the model, we released a large-scale, multi-modal 3D dance motion dataset, AIST++, which contains 5.2 hours of 3D dance motion in 1408 sequences, covering 10 dance genres, each including multi-view videos with known camera poses. Through extensive user studies on AIST++, we find that the FACT model outperforms recent state-of-the-art methods, both qualitatively and quantitatively.

We present a novel full-attention cross-modal transformer (FACT) network that can generate realistic 3D dance motion (right) conditioned on music and a new 3D dance dataset, AIST++ (left).

We generate the proposed 3D motion dataset from the existing AIST Dance Database — a collection of videos of dance with musical accompaniment, but without any 3D information. AIST contains 10 dance genres: Old School (Break, Pop, Lock and Waack) and New School (Middle Hip-Hop, LA-style Hip-Hop, House, Krump, Street Jazz and Ballet Jazz). Although it contains multi-view videos of dancers, these cameras are not calibrated.

For our purposes, we recovered the camera calibration parameters and the 3D human motion in terms of parameters used by the widely used SMPL 3D model. The resulting database, AIST++, is a large-scale, 3D human dance motion dataset that contains a wide variety of 3D motion, paired with music. Each frame includes extensive annotations:

• 9 views of camera intrinsic and extrinsic parameters;
• 17 COCO-format human joint locations in both 2D and 3D;
• 24 SMPL pose parameters along with the global scaling and translation.

The motions are equally distributed among all 10 dance genres, covering a wide variety of music tempos in beat per minute (BPM). Each genre of dance contains 85% basic movements and 15% advanced movements (longer choreographies freely designed by the dancers).

The AIST++ dataset also contains multi-view synchronized image data, making it useful for other research directions, such as 2D/3D pose estimation. To our knowledge, AIST++ is the largest 3D human dance dataset with 1408 sequences, 30 subjects and 10 dance genres, and with both basic and advanced choreographies.

An example of a 3D dance sequence in the AIST++ dataset. Left: Three views of the dance video from the AIST database. Right: Reconstructed 3D motion visualized in 3D mesh (top) and skeletons (bottom).

Because AIST is an instructional database, it records multiple dancers following the same choreography for different music with varying BPM, a common practice in dance. This posits a unique challenge in cross-modal sequence-to-sequence generation as the model needs to learn the one-to-many mapping between audio and motion. We carefully construct non-overlapping train and test subsets on AIST++ to ensure neither choreography nor music is shared across the subsets.

Full Attention Cross-Modal Transformer (FACT) Model
Using this data, we train the FACT model to generate 3D dance from music. The model begins by encoding seed motion and audio inputs using separate motion and audio transformers. The embeddings are then concatenated and sent to a cross-modal transformer, which learns the correspondence between both modalities and generates N future motion sequences. These sequences are then used to train the model in a self-supervised manner. All three transformers are jointly learned end-to-end. At test time, we apply this model in an autoregressive framework, where the predicted motion serves as the input to the next generation step. As a result, the FACT model is capable of generating long range dance motion frame-by-frame.

The FACT network takes in a music piece (Y) and a 2-second sequence of seed motion (X), then generates long-range future motions that correlate with the input music.

FACT involves three key design choices that are critical for producing realistic 3D dance motion from music.

1. All of the transformers use a full-attention mask, which can be more expressive than typical causal models because internal tokens have access to all inputs.
2. We train the model to predict N futures beyond the current input, instead of just the next motion. This encourages the network to pay more attention to the temporal context, and helps prevent the model from motion freezing or diverging after a few generation steps.
3. We fuse the two embeddings (motion and audio) early and employ a deep 12-layer cross-modal transformer module, which is essential for training a model that actually pays attention to the input music.

Results
We evaluate the performance based on three metrics:

Motion Quality: We calculate the Frechet Inception Distance (FID) between the real dance motion sequences in the AIST++ test set and 40 model generated motion sequences, each with 1200 frames (20 secs). We denote the FID based on the geometric and kinetic features as FID_g and FID_k, respectively.

Generation Diversity: Similar to prior work, to evaluate the model’s ability to generate divers dance motions, we calculate the average Euclidean distance in the feature space across 40 generated motions on the AIST++ test set, again comparing geometric feature space (Dist_g) and in the kinetic feature space (Dist_k).

Four different dance choreographies (right) generated using different music, but the same two second seed motion (left). The genres of the conditioning music are: Break, Ballet Jazz, Krump and Middle Hip-hop. The seed motion comes from hip-hop dance.

Motion-Music Correlation: Because there is no well-designed metric to measure the correlation between input music (music beats) and generated 3D motion (kinematic beats), we propose a novel metric, called Beat Alignment Score (BeatAlign).

Kinetic velocity (blue curve) and kinematic beats (green dotted line) of the generated dance motion, as well as the music beats (orange dotted line). The kinematic beats are extracted by finding local minima from the kinetic velocity curve.

Quantitative Evaluation
We compare the performance of FACT on each of these metrics to that of other state-of-the-art methods.

Compared to three recent state-of-the-art methods (Li et al., Dancenet, and Dance Revolution), the FACT model generates motions that are more realistic, better correlated with input music, and more diversified when conditioned on different music. *Note that the Li et al. generated motions are discontinuous, making the average kinetic feature distance abnormally high.

We also perceptually evaluate the motion-music correlation with a user study in which each participant is asked to watch 10 videos showing one of our results and one random counterpart, and then select which dancer is more in sync with the music. The study consisted of 30 participants, ranging from professional dancers to people who rarely dance. Compared to each baseline, 81% prefered the FACT model output to that of Li et al., 71% prefered FACT to Dancenet, and 77% prefered it Dance Revolution. Interestingly, 75% of participants preferred the unpaired AIST++ dance motion to that generated by FACT, which is unsurprising since the original dance captures are highly expressive.

Qualitative Results
Compared with prior methods like DanceNet (left) and Li et. al. (middle), 3D dance generated using the FACT model (right) is more realistic and better correlated with input music.

More generated 3D dances using the FACT model.

Conclusion and Discussion
We present a model that can not only learn the audio-motion correspondence, but also can generate high quality 3D motion sequences conditioned on music. Because generating 3D movement from music is a nascent area of study, we hope our work will pave the way for future cross-modal audio to 3D motion generation. We are also releasing AIST++, the largest 3D human dance dataset to date. This proposed, multi-view, multi-genre, cross-modal 3D motion dataset can not only help research in the conditional 3D motion generation research but also human understanding research in general. We are releasing the code in our GitHub repository and the trained model here.

While our results show a promising direction in this problem of music conditioned 3D motion generation, there are more to be explored. First, our approach is kinematic-based and we do not reason about physical interactions between the dancer and the floor. Therefore the global translation can lead to artifacts, such as foot sliding and floating. Second, our model is currently deterministic. Exploring how to generate multiple realistic dances per music is an exciting direction.

Acknowledgements
We gratefully acknowledge the contribution of other co-authors, including Ruilong Li and David Ross. We thank Chen Sun, Austin Myers, Bryan Seybold and Abhijit Kundu for helpful discussions. We thank Emre Aksan and Jiaman Li for sharing their code. We also thank Kevin Murphy for the early attempts in this direction, as well as Peggy Chi and Pan Chen for the help on user study experiments.

Posted by Katrin Tomanek, Software Engineer and Bob MacDonald, Technical Program Manager, Google Research

Speech impairments affect millions of people, with underlying causes ranging from neurological or genetic conditions to physical impairment, brain damage or hearing loss. Similarly, the resulting speech patterns are diverse, including stuttering, dysarthria, apraxia, etc., and can have a detrimental impact on self-expression, participation in society and access to voice-enabled technologies. Automatic speech recognition (ASR) technologies have the potential to help individuals with such speech impairments by improving access to dictation and home automation and by enhancing communication. However, while the increased computational power of deep learning systems and the availability of large training datasets has improved the accuracy of ASR systems, their performance is still insufficient for many people with speech disorders, rendering the technology unusable for many of the speakers who could benefit the most.

In 2019, we introduced Project Euphonia and discussed how we could use personalized ASR models of disordered speech to achieve accuracies on par with non-personalized ASR on typical speech. Today we share the results of two studies, presented at Interspeech 2021, that aim to expand the availability of personalized ASR models to more users. In “Disordered Speech Data Collection: Lessons Learned at 1 Million Utterances from Project Euphonia”, we present a greatly expanded collection of disordered speech data, composed of over 1 million utterances. Then, in “Automatic Speech Recognition of Disordered Speech: Personalized models outperforming human listeners on short phrases”, we discuss our efforts to generate personalized ASR models based on this corpus. This approach leads to highly accurate models that can achieve up to 85% improvement to the word error rate in select domains compared to out-of-the-box speech models trained on typical speech.

Impaired Speech Data Collection
Since 2019, speakers with speech impairments of varying degrees of severity across a variety of conditions have provided voice samples to support Project Euphonia’s research mission. This effort has grown Euphonia’s corpus to over 1 million utterances, comprising over 1400 hours from 1330 speakers (as of August 2021).

Distribution of severity of speech disorder and condition across all speakers with more than 300 utterances recorded. For conditions, only those with > 5 speakers are shown (all others aggregated into “OTHER” for k-anonymity). ALS = amyotrophic lateral sclerosis; DS = Down syndrome; PD = Parkinson’s disease; CP = cerebral palsy; HI = hearing impaired; MD = muscular dystrophy; MS = multiple sclerosis

To simplify the data collection, participants used an at-home recording system on their personal hardware (laptop or phone, with and without headphones), instead of an idealized lab-based setting that would collect studio quality recordings.

To reduce transcription cost, while still maintaining high transcript conformity, we prioritized scripted speech. Participants read prompts shown on a browser-based recording tool. Phrase prompts covered use-cases like home automation (“Turn on the TV.”), caregiver conversations (“I am hungry.”) and informal conversations (“How are you doing? Did you have a nice day?”). Most participants received a list of 1500 phrases, which included 1100 unique phrases along with 100 phrases that were each repeated four more times.

Speech professionals conducted a comprehensive auditory-perceptual speech assessment while listening to a subset of utterances for every speaker providing the following speaker-level metadata: speech disorder type (e.g., stuttering, dysarthria, apraxia), rating of 24 features of abnormal speech (e.g., hypernasality, articulatory imprecision, dysprosody), as well as recording quality assessments of both technical (e.g., signal dropouts, segmentation problems) and acoustic (e.g., environmental noise, secondary speaker crosstalk) features.

Personalized ASR Models
This expanded impaired speech dataset is the foundation of our new approach to personalized ASR models for disordered speech. Each personalized model uses a standard end-to-end, RNN-Transducer (RNN-T) ASR model that is fine-tuned using data from the target speaker only.

Architecture of RNN-Transducer. In our case, the encoder network consists of 8 layers and the predictor network consists of 2 layers of uni-directional LSTM cells.

To accomplish this, we focus on adapting the encoder network, i.e. the part of the model dealing with the specific acoustics of a given speaker, as speech sound disorders were most common in our corpus. We found that only updating the bottom five (out of eight) encoder layers while freezing the top three encoder layers (as well as the joint layer and decoder layers) led to the best results and effectively avoided overfitting. To make these models more robust against background noise and other acoustic effects, we employ a configuration of SpecAugment specifically tuned to the prevailing characteristics of disordered speech. Further, we found that the choice of the pre-trained base model was critical. A base model trained on a large and diverse corpus of typical speech (multiple domains and acoustic conditions) proved to work best for our scenario.

Results
We trained personalized ASR models for ~430 speakers who recorded at least 300 utterances. 10% of utterances were held out as a test set (with no phrase overlap) on which we calculated the word error rate (WER) for the personalized model and the unadapted base model.

Overall, our personalization approach yields significant improvements across all severity levels and conditions. Even for severely impaired speech, the median WER for short phrases from the home automation domain dropped from around 89% to 13%. Substantial accuracy improvements were also seen across other domains such as conversational and caregiver.

WER of unadapted and personalized ASR models on home automation phrases.

To understand when personalization does not work well, we analyzed several subgroups:

• HighWER and LowWER: Speakers with high and low personalized model WERs based on the 1st and 5th quintiles of the WER distribution.
• SurpHighWER: Speakers with a surprisingly high WER (participants with typical speech or mild speech impairment of the HighWER group).

Different pathologies and speech disorder presentations are expected to impact ASR non-uniformly. The distribution of speech disorder types within the HighWER group indicates that dysarthria due to cerebral palsy was particularly difficult to model. Not surprisingly, median severity was also higher in this group.

To identify the speaker-specific and technical factors that impact ASR accuracy, we examined the differences (Cohen’s D) in the metadata between the participants that had poor (HighWER) and excellent (LowWER) ASR performance. As expected, overall speech severity was significantly lower in the LowWER group than in the HighWER group (p < 0.01). Intelligibility and severity were the most prominent atypical speech features in the HighWER group; however, other speech features also emerged, including abnormal prosody, articulation, and phonation. These speech features are known to degrade overall speech intelligibility.

The SurpHighWER group had fewer training utterances and lower SNR compared with the LowWER group (p < 0.01) resulting in large (negative) effect sizes, with all other factors having small effect sizes, except fastness. In contrast, the HighWER group exhibited medium to large differences across all factors.

Speech disorder and technical metadata effect sizes for the HighWER-vs-LowWER and SurpHighWER-vs-LowWER pairs. Positive effects indicated that the group values of the HighWER group were greater than LowWER groups.

We then compared personalized ASR models to human listeners. Three speech professionals independently transcribed 30 utterances per speaker. We found that WERs were, on average, lower for personalized ASR models compared to the WERs of human listeners, with gains increasing by severity.

Delta between the WERs of the personalized ASR models and the human listeners. Negative values indicate that personalized ASR performs better than human (expert) listeners.

Conclusions
With over 1 million utterances, Euphonia’s corpus is one of the largest and most diversely disordered speech corpora (in terms of disorder types and severities) and has enabled significant advances in ASR accuracy for these types of atypical speech. Our results demonstrate the efficacy of personalized ASR models for recognizing a wide range of speech impairments and severities, with potential for making ASR available to a wider population of users.

Acknowledgements
Key contributors to this project include Michael Brenner, Julie Cattiau, Richard Cave, Jordan Green, Rus Heywood, Pan-Pan Jiang, Anton Kast, Marilyn Ladewig, Bob MacDonald, Phil Nelson, Katie Seaver, Jimmy Tobin, and Katrin Tomanek. We gratefully acknowledge the support Project Euphonia received from members of many speech research teams across Google, including Françoise Beaufays, Fadi Biadsy, Dotan Emanuel, Khe Chai Sim, Pedro Moreno Mengibar, Arun Narayanan, Hasim Sak, Suzan Schwartz, Joel Shor, and many others. And most importantly, we wanted to say a huge thank you to the over 1300 participants who recorded speech samples and the many advocacy groups who helped us connect with these participants.