As wearables and handheld devices decrease in size, haptics become an increasingly vital channel for feedback, be it through silent alerts or a subtle "click" sensation when pressing buttons on a touch screen. Haptic feedback, ubiquitous in nearly all wearable devices and mobile phones, is commonly enabled by a linear resonant actuator (LRA), a small linear motor that leverages resonance to provide a strong haptic signal in a small package. However, the touch and pressure sensing needed to activate the haptic feedback tend to depend on additional, separate hardware which increases the price, size and complexity of the system.
In “Haptics with Input: Back-EMF in Linear Resonant Actuators to Enable Touch, Pressure and Environmental Awareness”, presented at ACM UIST 2020, we demonstrate that widely available LRAs can sense a wide range of external information, such as touch, tap and pressure, in addition to being able to relay information about contact with the skin, objects and surfaces. We achieve this with off-the-shelf LRAs by multiplexing the actuation with short pulses of custom waveforms that are designed to enable sensing using the back-EMF voltage. We demonstrate the potential of this approach to enable expressive discrete buttons and vibrotactile interfaces and show how the approach could bring rich sensing opportunities to integrated haptics modules in mobile devices, increasing sensing capabilities with fewer components. Our technique is potentially compatible with many existing LRA drivers, as they already employ back-EMF sensing for autotuning of the vibration frequency.
Back-EMF Principle in an LRA Inside the LRA enclosure is a magnet attached to a small mass, both moving freely on a spring. The magnet moves in response to the excitation voltage introduced by the voice coil. The motion of the oscillating mass produces a counter-electromotive force, or back-EMF, which is a voltage proportional to the rate of change of magnetic flux. A greater oscillation speed creates a larger back-EMF voltage, while a stationary mass generates zero back-EMF voltage.
Active Back-EMF for Sensing Touching or making contact with the LRA during vibration changes the velocity of the interior mass, as energy is dissipated into the contact object. This works well with soft materials that deform under pressure, such as the human body. A finger, for example, absorbs different amounts of energy depending on the contact force as it flattens against the LRA. By driving the LRA with small amounts of energy, we can measure this phenomenon using the back-EMF voltage. Because leveraging the back-EMF behavior for sensing is an active process, the key insight that enabled this work was the design of a custom, off-resonance driver waveform that allows continuous sensing while minimizing vibrations, sound and power consumption.
We measure back-EMF from the floating voltage between the two LRA leads, which requires disconnecting the motor driver briefly to avoid disturbances. While the driver is disconnected, the mass is still oscillating inside the LRA, producing an oscillating back-EMF voltage. Because commercial back-EMF sensing LRA drivers do not provide the raw data, we designed a custom circuit that is able to pick up and amplify small back-EMF voltage. We also generated custom drive pulses that minimize vibrations and energy consumption.
Applications The behavior of the LRAs used in mobile phones is the same, whether they are on a table, on a soft surface, or hand held. This may cause problems, as a vibrating phone could slide off a glass table or emit loud and unnecessary vibrating sounds. Ideally, the LRA on a phone would automatically adjust based on its environment. We demonstrate our approach for sensing using the LRA back-EMF technique by wiring directly to a Pixel 4’s LRA, and then classifying whether the phone is held in hand, placed on a soft surface (foam), or placed on a table.
We also present a prototype that demonstrates how LRAs could be used as combined input/output devices in portable electronics. We attached two LRAs, one on the left and one on the right side of a phone. The buttons provide tap, touch, and pressure sensing. They are also programmed to provide haptic feedback, once the touch is detected.
There are a number of wearable tactile aid devices, such as sleeves, vests, and bracelets. To transmit tactile feedback to the skin with consistent force, the tactor has to apply the right pressure; it can not be too loose or too tight. Currently, the typical way to do so is through manual adjustment, which can be inconsistent and lacks measurable feedback. We show how the LRA back-EMF technique can be used to continuously monitor the fit bracelet device and prompt the user if it's too tight, too loose, or just right.
Evaluating an LRA as a Sensor The LRA works well as a pressure sensor, because it has a quadratic response to the force magnitude during touch. Our method works for all five off-the-shelf LRA types that we evaluated. Because the typical power consumption is only 4.27 mA, all-day sensing would only reduce the battery life of a Pixel 4 phone from 25 to 24 hours. The power consumption can be greatly reduced by using low-power amplifiers and employing active sensing only when needed, such as when the phone is active and interacting with the user.
The challenge with active sensing is to minimize vibrations, so they are not perceptible when touching and do not produce audible sound. We optimize the active sensing to produce only 2 dB of sound and 0.45 m/s2 peak-to-peak acceleration, which is just barely perceptible by finger and is quiet, in contrast to the regular 8.49 m/s2 .
Future Work and Conclusion To see the work presented here in action, please see the video below.
In the future, we plan to explore other sensing techniques, perhaps measuring the current could be an alternative approach. Also, using machine learning could potentially improve the sensing and provide more accurate classification of the complex back-EMF patterns. Our method could be developed further to enable closed-loop feedback with the actuator and sensor, which would allow the actuator to provide the same force, regardless of external conditions.
We believe that this work opens up new opportunities for leveraging existing ubiquitous hardware to provide rich interactions and closed-loop feedback haptic actuators.
Acknowledgments This work was done by Artem Dementyev, Alex Olwal, and Richard Lyon. Thanks to Mathieu Le Goc and Thad Starner for feedback on the paper.
Creating art for digital video games takes a high degree of artistic creativity and technical knowledge, while also requiring game artists to quickly iterate on ideas and produce a high volume of assets, often in the face of tight deadlines. What if artists had a paintbrush that acted less like a tool and more like an assistant? A machine learning model acting as such a paintbrush could reduce the amount of time necessary to create high-quality art without sacrificing artistic choices, perhaps even enhancing creativity.
Today, we present Chimera Painter, a trained machine learning (ML) model that automatically creates a fully fleshed out rendering from a user-supplied creature outline. Employed as a demo application, Chimera Painter adds features and textures to a creature outline segmented with body part labels, such as “wings” or “claws”, when the user clicks the “transform” button. Below is an example using the demo with one of the preset creature outlines.
In this post, we describe some of the challenges in creating the ML model behind Chimera Painter and demonstrate how one might use the tool for the creation of video game-ready assets.
Prototyping for a New Type of Model In developing an ML model to produce video-game ready creature images, we created a digital card game prototype around the concept of combining creatures into new hybrids that can then battle each other. In this game, a player would begin with cards of real-world animals (e.g., an axolotl or a whale) and could make them more powerful by combining them (making the dreaded Axolotl-Whale chimera). This provided a creative environment for demonstrating an image-generating model, as the number of possible chimeras necessitated a method for quickly designing large volumes of artistic assets that could be combined naturally, while still retaining identifiable visual characteristics of the original creatures.
Since our goal was to create high-quality creature card images guided by artist input, we experimented with generative adversarial networks (GANs), informed by artist feedback, to create creature images that would be appropriate for our fantasy card game prototype. GANs pair two convolutional neural networks against each other: a generator network to create new images and a discriminator network to determine if these images are samples from the training dataset (in this case, artist-created images) or not. We used a variant called a conditional GAN, where the generator takes a separate input to guide the image generation process. Interestingly, our approach was a strict departure from other GAN efforts, which typically focus on photorealism.
To train the GANs, we created a dataset of full color images with single-species creature outlines adapted from 3D creature models. The creature outlines characterized the shape and size of each creature, and provided a segmentation map that identified individual body parts. After model training, the model was tasked with generating multi-species chimeras, based on outlines provided by artists. The best performing model was then incorporated into Chimera Painter. Below we show some sample assets generated using the model, including single-species creatures, as well as the more complex multi-species chimeras.
Learning to Generate Creatures with Structure An issue with using GANs for generating creatures was the potential for loss of anatomical and spatial coherence when rendering subtle or low-contrast parts of images, despite these being of high perceptual importance to humans. Examples of this can include eyes, fingers, or even distinguishing between overlapping body parts with similar textures (see the affectionately named BoggleDog below).
Generating chimeras required a new non-photographic fantasy-styled dataset with unique characteristics, such as dramatic perspective, composition, and lighting. Existing repositories of illustrations were not appropriate to use as datasets for training an ML model, because they may be subject to licensing restrictions, have conflicting styles, or simply lack the variety needed for this task.
To solve this, we developed a new artist-led, semi-automated approach for creating an ML training dataset from 3D creature models, which allowed us to work at scale and rapidly iterate as needed. In this process, artists would create or obtain a set of 3D creature models, one for each creature type needed (such as hyenas or lions). Artists then produced two sets of textures that were overlaid on the 3D model using the Unreal Engine — one with the full color texture (left image, below) and the other with flat colors for each body part (e.g., head, ears, neck, etc), called a “segmentation map” (right image, below). This second set of body part segments was given to the model at training to ensure that the GAN learned about body part-specific structure, shapes, textures, and proportions for a variety of creatures.
The 3D creature models were all placed in a simple 3D scene, again using the Unreal Engine. A set of automated scripts would then take this 3D scene and interpolate between different poses, viewpoints, and zoom levels for each of the 3D creature models, creating the full color images and segmentation maps that formed the training dataset for the GAN. Using this approach, we generated 10,000+ image + segmentation map pairs per 3D creature model, saving the artists millions of hours of time compared to creating such data manually (at approximately 20 minutes per image).
Fine Tuning The GAN had many different hyper-parameters that could be adjusted, leading to different qualities in the output images. In order to better understand which versions of the model were better than others, artists were provided samples for different creature types generated by these models and asked to cull them down to a few best examples. We gathered feedback about desired characteristics present in these examples, such as a feeling of depth, style with regard to creature textures, and realism of faces and eyes. This information was used both to train new versions of the model and, after the model had generated hundreds of thousands of creature images, to select the best image from each creature category (e.g., gazelle, lynx, gorilla, etc).
We tuned the GAN for this task by focusing on the perceptual loss. This loss function component (also used in Stadia’s Style Transfer ML) computes a difference between two images using extracted features from a separate convolutional neural network (CNN) that was previously trained on millions of photographs from the ImageNet dataset. The features are extracted from different layers of the CNN and a weight is applied to each, which affects their contribution to the final loss value. We discovered that these weights were critically important in determining what a final generated image would look like. Below are some examples from the GAN trained with different perceptual loss weights.
Some of the variation in the images above is due to the fact that the dataset includes multiple textures for each creature (for example, a reddish or grayish version of the bat). However, ignoring the coloration, many differences are directly tied to changes in perceptual loss values. In particular, we found that certain values brought out sharper facial features (e.g., bottom right vs. top right) or “smooth” versus “patterned” (top right vs. bottom left) that made generated creatures feel more real.
Here are some creatures generated from the GAN trained with different perceptual loss weights, showing off a small sample of the outputs and poses that the model can handle.
Chimera Painter The trained GAN is now available in the Chimera Painter demo, allowing artists to work iteratively with the model, rather than drawing dozens of similar creatures from scratch. An artist can select a starting point and then adjust the shape, type, or placement of creature parts, enabling rapid exploration and for the creation of a large volume of images. The demo also allows for uploading a creature outline created in an external program, like Photoshop. Simply download one of the preset creature outlines to get the colors needed for each creature part and use this as a template for drawing one outside of Chimera Painter, and then use the “Load’ button on the demo to use this outline to flesh out your creation.
It is our hope that these GAN models and the Chimera Painter demonstration tool might inspire others to think differently about their art pipeline. What can one create when using machine learning as a paintbrush?
Acknowledgments This project is conducted in collaboration with many people. Thanks to Ryan Poplin, Lee Dotson, Trung Le, Monica Dinculescu, Marc Destefano, Aaron Cammarata, Maggie Oh, Richard Wu, Ji Hun Kim, Erin Hoffman-John, and Colin Boswell. Thanks to everyone who pitched in to give hours of art direction, technical feedback, and drawings of fantastic creatures.
The responsible research and development of machine learning (ML) can play a pivotal role in helping to solve a wide variety of societal challenges. At Google, our research reflects our AI Principles, from helping to protect patients from medication errors and improving flood forecasting models, to presenting methods that tackle unfair bias in products, such as Google Translate, and providing resources for other researchers to do the same.
One broad category for applying ML responsibly is the task of classification — systems that sort data into labeled categories. At Google, such models are used throughout our products to enforce policies, ranging from the detection of hate speech to age-appropriate content filtering. While these classifiers serve vital functions, it is also essential that they are built in ways that minimize unfair biases for users.
Today, we are announcing the release of MinDiff, a new regularization technique available in the TF Model Remediation library for effectively and efficiently mitigating unfair biases when training ML models. In this post, we discuss the research behind this technique and explain how it addresses the practical constraints and requirements we’ve observed when incorporating it in Google’s products.
Unfair Biases in Classifiers To illustrate how MinDiff can be used, consider an example of a product policy classifier that is tasked with identifying and removing text comments that could be considered toxic. One challenge is to make sure that the classifier is not unfairly biased against submissions from a particular group of users, which could result in incorrect removal of content from these groups.
The academic community has laid a solid theoretical foundation for ML fairness, offering a breadth of perspectives on what unfair bias means and on the tensions between different frameworks for evaluating fairness. One of the most common metrics is equality of opportunity, which, in our example, means measuring and seeking to minimize the difference in false positive rate (FPR) across groups. In the example above, this means that the classifier should not be more likely to incorrectly remove safe comments from one group than another. Similarly, the classifier’s false negative rate should be equal between groups. That is, the classifier should not miss toxic comments against one group more than it does for another.
When the end goal is to improve products, it’s important to be able to scale unfair bias mitigation to many models. However, this poses a number of challenges:
MinDiff Framework We iteratively developed the MinDiff framework over the previous few years to meet these design requirements. Because demographic information is so rarely known, we utilize in-process approaches in which the model’s training objective is augmented with an objective specifically focused on removing biases. This new objective is then optimized over the small sample of data with known demographic information. To improve ease of use, we switched from adversarial training to a regularization framework, which penalizes statistical dependency between its predictions and demographic information for non-harmful examples. This encourages the model to equalize error rates across groups, e.g., classifying non-harmful examples as toxic.
There are several ways to encode this dependency between predictions and demographic information. Our initial MinDiff implementation minimized the correlation between the predictions and the demographic group, which essentially optimized for the average and variance of predictions to be equal across groups, even if the distributions still differ afterward. We have since improved MinDiff further by considering the maximum mean discrepancy (MMD) loss, which is closer to optimizing for the distribution of predictions to be independent of demographics. We have found that this approach is better able to both remove biases and maintain model accuracy.
To date we have launched modeling improvements across several classifiers at Google that moderate content quality. We went through multiple iterations to develop a robust, responsible, and scalable approach, solving research challenges and enabling broad adoption.
Gaps in error rates of classifiers is an important set of unfair biases to address, but not the only one that arises in ML applications. For ML researchers and practitioners, we hope this work can further advance research toward addressing even broader classes of unfair biases and the development of approaches that can be used in practical applications. In addition, we hope that the release of the MinDiff library and the associated demos and documentation, along with the tools and experience shared here, can help practitioners improve their models and products.
Acknowledgements This research effort on ML Fairness in classification was jointly led with Jilin Chen, Shuo Chen, Ed H. Chi, Tulsee Doshi, and Hai Qian. Further, this work was pursued in collaboration with Jonathan Bischof, Qiuwen Chen, Pierre Kreitmann, and Christine Luu. The MinDiff infrastructure was also developed in collaboration with Nick Blumm, James Chen, Thomas Greenspan, Christina Greer, Lichan Hong, Manasi Joshi, Maciej Kula, Summer Misherghi, Dan Nanas, Sean O'Keefe, Mahesh Sathiamoorthy, Catherina Xu, and Zhe Zhao. (All names are listed in alphabetical order of last names.)
Melodies stuck in your head, often referred to as “earworms,” are a well-known and sometimes irritating phenomenon — once that earworm is there, it can be tough to get rid of it. Research has found that engaging with the original song, whether that’s listening to or singing it, will drive the earworm away. But what if you can’t quite recall the name of the song, and can only hum the melody?
Existing methods to match a hummed melody to its original polyphonic studio recording face several challenges. With lyrics, background vocals and instruments, the audio of a musical or studio recording can be quite different from a hummed tune. By mistake or design, when someone hums their interpretation of a song, often the pitch, key, tempo or rhythm may vary slightly or even significantly. That’s why so many existing approaches to query by humming match the hummed tune against a database of pre-existing melody-only or hummed versions of a song, instead of identifying the song directly. However, this type of approach often relies on a limited database that requires manual updates.
Launched in October, Hum to Search is a new fully machine-learned system within Google Search that allows a person to find a song using only a hummed rendition of it. In contrast to existing methods, this approach produces an embedding of a melody from a spectrogram of a song without generating an intermediate representation. This enables the model to match a hummed melody directly to the original (polyphonic) recordings without the need for a hummed or MIDI version of each track or for other complex hand-engineered logic to extract the melody. This approach greatly simplifies the database for Hum to Search, allowing it to constantly be refreshed with embeddings of original recordings from across the world — even the latest releases.
Background Many existing music recognition systems convert an audio sample into a spectrogram before processing it, in order to find a good match. However, one challenge in recognizing a hummed melody is that a hummed tune often contains relatively little information, as illustrated by this hummed example of Bella Ciao. The difference between the hummed version and the same segment from the corresponding studio recording can be visualized using spectrograms, seen below:
Given the image on the left, a model needs to locate the audio corresponding to the right-hand image from a collection of over 50M similar-looking images (corresponding to segments of studio recordings of other songs). To achieve this, the model has to learn to focus on the dominant melody, and ignore background vocals, instruments, and voice timbre, as well as differences stemming from background noise or room reverberations. To find by eye the dominant melody that might be used to match these two spectrograms, a person might look for similarities in the lines near the bottom of the above images.
Prior efforts to enable discovery of music, in particular in the context of recognizing recorded music being played in an environment such as a cafe or a club, demonstrated how machine learning might be applied to this problem. Now Playing, released to Pixel phones in 2017, uses an on-device deep neural network to recognize songs without the need for a server connection, and Sound Search further developed this technology to provide a server-based recognition service for faster and more accurate searching of over 100 million songs. The next challenge then was to leverage what was learned from these releases to recognize hummed or sung music from a similarly large library of songs.
Machine Learning Setup The first step in developing Hum to Search was to modify the music-recognition models used in Now Playing and Sound Search to work with hummed recordings. In principle, many such retrieval systems (e.g., image recognition) work in a similar way. A neural network is trained with pairs of input (here pairs of hummed or sung audio with recorded audio) to produce embeddings for each input, which will later be used for matching to a hummed melody.
To enable humming recognition, the network should produce embeddings for which pairs of audio containing the same melody are close to each other, even if they have different instrumental accompaniment and singing voices. Pairs of audio containing different melodies should be far apart. In training, the network is provided such pairs of audio until it learns to produce embeddings with this property.
The trained model can then generate an embedding for a tune that is similar to the embedding of the song’s reference recording. Finding the correct song is then only a matter of searching for similar embeddings from a database of reference recordings computed from audio of popular music.
Training Data Because training of the model required song pairs (recorded and sung), the first challenge was to obtain enough training data. Our initial dataset consisted of mostly sung music segments (very few of these contained humming). To make the model more robust, we augmented the audio during training, for example by varying the pitch or tempo of the sung input randomly. The resulting model worked well enough for people singing, but not for people humming or whistling.
To improve the model’s performance on hummed melodies we generated additional training data of simulated “hummed” melodies from the existing audio dataset using SPICE, a pitch extraction model developed by our wider team as part of the FreddieMeter project. SPICE extracts the pitch values from given audio, which we then use to generate a melody consisting of discrete audio tones. The very first version of this system transformed this original clip into these tones.
We later refined this approach by replacing the simple tone generator with a neural network that generates audio resembling an actual hummed or whistled tune. For example, the network generates this humming example or whistling example from the above sung clip.
As a final step, we compared training data by mixing and matching the audio samples. For example, if we had a similar clip from two different singers, we’d align those two clips with our preliminary models, and are therefore able to show the model an additional pair of audio clips that represent the same melody.
Machine Learning Improvements When training the Hum to Search model, we started with a triplet loss function. This loss has been shown to perform well across a variety of classification tasks like images and recorded music. Given a pair of audio corresponding to the same melody (points R and P in embedding space, shown below), triplet loss would ignore certain parts of the training data derived from a different melody. This helps the machine improve learning behavior, either when it finds a different melody that is too ‘easy’ in that it is already far away from R and P (see point E) or because it is too hard in that, given the model's current state of learning, the audio ends up being too close to R — even though according to our data it represents a different melody (see point H).
We’ve found that we could improve the accuracy of the model by taking these additional training data (points H and E) into account, namely by formulating a general notion of model confidence across a batch of examples: How sure is the model that all the data it has seen can be classified correctly, or has it seen examples that do not fit its current understanding? Based on this notion of confidence, we added a loss that drives model confidence towards 100% across all areas of the embedding space, which led to improvements in our model’s precision and recall.
The above changes, but in particular our variations, augmentations and superpositions of the training data, enabled the neural network model deployed in Google Search to recognize sung or hummed melodies. The current system reaches a high level of accuracy on a song database that contains over half a million songs that we are continually updating. This song corpus still has room to grow to include more of the world’s many melodies.
To try the feature, you can open the latest version of the Google app, tap the mic icon and say “what's this song?” or click the “Search a song” button, after which you can hum, sing, or whistle away! We hope that Hum to Search can help with that earworm of yours, or maybe just help you in case you want to find and playback a song without having to type its name.
Acknowledgements The work described here was authored by Alex Tudor, Duc Dung Nguyen, Matej Kastelic, Mihajlo Velimirović, Stefan Christoph, Mauricio Zuluaga, Christian Frank, Dominik Roblek, and Matt Sharifi. We would like to deeply thank Krishna Kumar, Satyajeet Salgar and Blaise Aguera y Arcas for their ongoing support, as well as all the Google teams we've collaborated with to build the full Hum to Search product.
We would also like to thank all our colleagues at Google who donated clips of themselves singing or humming and therefore laid a foundation for this work, as well as Nick Moukhine for building the Google-internal singing donation app. Finally, special thanks to Meghan Danks and Krishna Kumar for their feedback on earlier versions of this post.
Voice assistive technologies, which enable users to employ voice commands to interact with their devices, rely on accurate speech recognition to ensure responsiveness to a specific user. But in many real-world use cases, the input to such technologies often consists of overlapping speech, which poses great challenges to many speech recognition algorithms. In 2018, we published a VoiceFilter system, which leverages Google’s Voice Match to personalize interaction with assistive technology by allowing people to enroll their voices.
While the VoiceFilter approach is highly successful, achieving a better source to distortion ratio (SDR) than conventional approaches, efficient on-device streaming speech recognition requires addressing restrictions such as model size, CPU and memory limitations, as well as battery usage considerations and latency minimization.
In “VoiceFilter-Lite: Streaming Targeted Voice Separation for On-Device Speech Recognition”, we present an update to VoiceFilter for on-device use that can significantly improve speech recognition in overlapping speech by leveraging the enrolled voice of a selected speaker. Importantly, this model can be easily integrated with existing on-device speech recognition applications, allowing the user to access voice assistive features under extremely noisy conditions even if an internet connection is unavailable. Our experiments show that a 2.2MB VoiceFilter-Lite model provides a 25.1% improvement to the word error rate (WER) on overlapping speech.
Improving On-Device Speech Recognition While the original VoiceFilter system was very successful at separating a target speaker's speech signal from other overlapping sources, its model size, computational cost and latency are not feasible for speech recognition on mobile devices.
The new VoiceFilter-Lite system has been carefully designed to fit on-device applications. Instead of processing audio waveforms, VoiceFilter-Lite takes exactly the same input features as the speech recognition model (stacked log Mel-filterbanks), and directly enhances these features by filtering out components not belonging to the target speaker in real time. Together with several optimizations on network topologies, the number of runtime operations is drastically reduced. After quantizing the neural network with the TensorFlow Lite library, the model size is only 2.2 MB, which fits most on-device applications.
To train the VoiceFilter-Lite model, the filterbanks of the noisy speech are fed as input to the network together with an embedding vector that represents the identity of the target speaker (i.e., a d-vector). The network predicts a mask that is element-wise multiplied to the input to produce enhanced filterbanks. A loss function is defined to minimize the difference between the enhanced filterbanks and the filterbanks from the clean speech during training.
VoiceFilter-Lite is a plug-and-play model, which allows the application in which it’s implemented to easily bypass it if the speaker did not enroll their voice. This also means that the speech recognition model and the VoiceFilter-Lite model can be separately trained and updated, which largely reduces engineering complexity in the deployment process.
Addressing the Challenge of Over-Suppression When speech separation models are used for improving speech recognition, two types of error could occur: under-suppression, when the model fails to filter out noisy components from the signal; and over-suppression, when the model fails to preserve useful signal, resulting in some words being dropped from the recognized text. Over-suppression is especially problematic since modern speech recognition models are usually already trained with extensively augmented data (such as room simulation and SpecAugment), and thus are more robust to under-suppression.
VoiceFilter-Lite addresses the over-suppression issue with two novel approaches. First, it uses an asymmetric loss during the training process, such that the model is less tolerant to over-suppression than under-suppression. Second, it predicts the type of noise at runtime, and adaptively adjusts the suppression strength according to this prediction.
With these two solutions, the VoiceFilter-Lite model retains great performance on streaming speech recognition for other scenarios, such as single-speaker speech under quiet or various noise conditions, while still providing significant improvement on overlapping speech. From our experiments, we observed a 25.1% improvement of word error rate after the 2.2MB VoiceFilter-Lite model is applied on additive overlapping speech. For reverberant overlapping speech, which is a more challenging task to simulate far-field devices such as smart home speakers, we also observed a 14.7% improvement of word error rate with VoiceFilter-Lite.
Future Work While VoiceFilter-Lite has shown great promise for various on-device speech applications, we are also exploring several other directions to make VoiceFilter-Lite more useful. First, our current model is trained and evaluated with English speech only. We are excited about adopting the same technology to improve speech recognition for more languages. Second, we would like to directly optimize the speech recognition loss during the training of VoiceFilter-Lite, which can potentially further improve speech recognition beyond overlapping speech.
Acknowledgements The research described in this post represents joint efforts from multiple teams within Google. Contributors include Quan Wang, Ignacio Lopez Moreno, Mert Saglam, Kevin Wilson, Alan Chiao, Renjie Liu, Yanzhang He, Wei Li, Jason Pelecanos, Philip Chao, Sinan Akay, John Han, Stephen Wu, Hannah Muckenhirn, Ye Jia, Zelin Wu, Yiteng Huang, Marily Nika, Jaclyn Konzelmann, Nino Tasca, and Alexander Gruenstein.
The state of the art in machine learning (ML) has achieved exceptional accuracy on many computer vision tasks solely by training models on photos. Building upon these successes and advancing 3D object understanding has great potential to power a wider range of applications, such as augmented reality, robotics, autonomy, and image retrieval. For example, earlier this year we released MediaPipe Objectron, a set of real-time 3D object detection models designed for mobile devices, which were trained on a fully annotated, real-world 3D dataset, that can predict objects’ 3D bounding boxes.
Yet, understanding objects in 3D remains a challenging task due to the lack of large real-world datasets compared to 2D tasks (e.g., ImageNet, COCO, and Open Images). To empower the research community for continued advancement in 3D object understanding, there is a strong need for the release of object-centric video datasets, which capture more of the 3D structure of an object, while matching the data format used for many vision tasks (i.e., video or camera streams), to aid in the training and benchmarking of machine learning models.
Today, we are excited to release the Objectron dataset, a collection of short, object-centric video clips capturing a larger set of common objects from different angles. Each video clip is accompanied by AR session metadata that includes camera poses and sparse point-clouds. The data also contain manually annotated 3D bounding boxes for each object, which describe the object’s position, orientation, and dimensions. The dataset consists of 15K annotated video clips supplemented with over 4M annotated images collected from a geo-diverse sample (covering 10 countries across five continents).
A 3D Object Detection Solution Along with the dataset, we are also sharing a 3D object detection solution for four categories of objects — shoes, chairs, mugs, and cameras. These models are released in MediaPipe, Google's open source framework for cross-platform customizable ML solutions for live and streaming media, which also powers ML solutions like on-device real-time hand, iris and body pose tracking.
In contrast to the previously released single-stage Objectron model, these newest versions utilize a two-stage architecture. The first stage employs the TensorFlow Object Detection model to find the 2D crop of the object. The second stage then uses the image crop to estimate the 3D bounding box while simultaneously computing the 2D crop of the object for the next frame, so that the object detector does not need to run every frame. The second stage 3D bounding box predictor runs at 83 FPS on Adreno 650 mobile GPU.
Evaluation Metric for 3D Object Detection With ground truth annotations, we evaluate the performance of 3D object detection models using 3D intersection over union (IoU) similarity statistics, a commonly used metric for computer vision tasks, which measures how close the bounding boxes are to the ground truth.
We propose an algorithm for computing accurate 3D IoU values for general 3D-oriented boxes. First, we compute the intersection points between faces of the two boxes using Sutherland-Hodgman Polygon clipping algorithm. This is similar to frustum culling, a technique used in computer graphics. The volume of the intersection is computed by the convex hull of all the clipped polygons. Finally, the IoU is computed from the volume of the intersection and volume of the union of two boxes. We are releasing the evaluation metrics source code along with the dataset.
Dataset Format The technical details of the Objectron dataset, including usage and tutorials, are available on the dataset website. The dataset includes bikes, books, bottles, cameras, cereal boxes, chairs, cups, laptops, and shoes, and is stored in the objectron bucket on Google Cloud storage with the following assets:
With the dataset, we are also open-sourcing a data-pipeline to parse the dataset in popular Tensorflow, PyTorch and Jax frameworks. Example colab notebooks are also provided.
By releasing this Objectron dataset, we hope to enable the research community to push the limits of 3D object geometry understanding. We also hope to foster new research and applications, such as view synthesis, improved 3D representation, and unsupervised learning. Stay tuned for future activities and developments by joining our mailing list and visiting our github page.
Acknowledgements The research described in this post was done by Adel Ahmadyan, Liangkai Zhang, Jianing Wei, Artsiom Ablavatski, Mogan Shieh, Ryan Hickman, Buck Bourdon, Alexander Kanaukou, Chuo-Ling Chang, Matthias Grundmann, and Tom Funkhouser. We thank Aliaksandr Shyrokau, Sviatlana Mialik, Anna Eliseeva, and the annotation team for their high quality annotations. We also would like to thank Jonathan Huang and Vivek Rathod for their guidance on TensorFlow Object Detection API.
Video conferencing is becoming ever more critical in people's work and personal lives. Improving that experience with privacy enhancements or fun visual touches can help center our focus on the meeting itself. As part of this goal, we recently announced ways to blur and replace your background in Google Meet, which use machine learning (ML) to better highlight participants regardless of their surroundings. Whereas other solutions require installing additional software, Meet’s features are powered by cutting-edge web ML technologies built with MediaPipe that work directly in your browser — no extra steps necessary. One key goal in developing these features was to provide real-time, in-browser performance on almost all modern devices, which we accomplished by combining efficient on-device ML models, WebGL-based rendering, and web-based ML inference via XNNPACK and TFLite.
Overview of Our Web ML Solution The new features in Meet are developed with MediaPipe, Google's open source framework for cross-platform customizable ML solutions for live and streaming media, which also powers ML solutions like on-device real-time hand, iris and body pose tracking.
A core need for any on-device solution is to achieve high performance. To accomplish this, MediaPipe’s web pipeline leverages WebAssembly, a low-level binary code format designed specifically for web browsers that improves speed for compute-heavy tasks. At runtime, the browser converts WebAssembly instructions into native machine code that executes much faster than traditional JavaScript code. In addition, Chrome 84 recently introduced support for WebAssembly SIMD, which processes multiple data points with each instruction, resulting in a performance boost of more than 2x.
Our solution first processes each video frame by segmenting a user from their background (more about our segmentation model later in the post) utilizing ML inference to compute a low resolution mask. Optionally, we further refine the mask to align it with the image boundaries. The mask is then used to render the video output via WebGL2, with the background blurred or replaced.
In the current version, model inference is executed on the client’s CPU for low power consumption and widest device coverage. To achieve real-time performance, we designed efficient ML models with inference accelerated by the XNNPACK library, the first inference engine specifically designed for the novel WebAssembly SIMD specification. Accelerated by XNNPACK and SIMD, the segmentation model can run in real-time on the web.
Enabled by MediaPipe's flexible configuration, the background blur/replace solution adapts its processing based on device capability. On high-end devices it runs the full pipeline to deliver the highest visual quality, whereas on low-end devices it continues to perform at speed by switching to compute-light ML models and bypassing the mask refinement.
Segmentation Model On-device ML models need to be ultra lightweight for fast inference, low power consumption, and small download size. For models running in the browser, the input resolution greatly affects the number of floating-point operations (FLOPs) necessary to process each frame, and therefore needs to be small as well. We downsample the image to a smaller size before feeding it to the model. Recovering a segmentation mask as fine as possible from a low-resolution image adds to the challenges of model design.
The overall segmentation network has a symmetric structure with respect to encoding and decoding, while the decoder blocks (light green) also share a symmetric layer structure with the encoder blocks (light blue). Specifically, channel-wise attention with global average pooling is applied in both encoder and decoder blocks, which is friendly to efficient CPU inference.
We modified MobileNetV3-small as the encoder, which has been tuned by network architecture search for the best performance with low resource requirements. To reduce the model size by 50%, we exported our model to TFLite using float16 quantization, resulting in a slight loss in weight precision but with no noticeable effect on quality. The resulting model has 193K parameters and is only 400KB in size.
Rendering Effects Once segmentation is complete, we use OpenGL shaders for video processing and effect rendering, where the challenge is to render efficiently without introducing artifacts. In the refinement stage, we apply a joint bilateral filter to smooth the low resolution mask.
The blur shader simulates a bokeh effect by adjusting the blur strength at each pixel proportionally to the segmentation mask values, similar to the circle-of-confusion (CoC) in optics. Pixels are weighted by their CoC radii, so that foreground pixels will not bleed into the background. We implemented separable filters for the weighted blur, instead of the popular Gaussian pyramid, as it removes halo artifacts surrounding the person. The blur is performed at a low resolution for efficiency, and blended with the input frame at the original resolution.
For background replacement, we adopt a compositing technique, known as light wrapping, for blending segmented persons and customized background images. Light wrapping helps soften segmentation edges by allowing background light to spill over onto foreground elements, making the compositing more immersive. It also helps minimize halo artifacts when there is a large contrast between the foreground and the replaced background.
Performance To optimize the experience for different devices, we provide model variants at multiple input sizes (i.e., 256x144 and 160x96 in the current release), automatically selecting the best according to available hardware resources.
We evaluated the speed of model inference and the end-to-end pipeline on two common devices: MacBook Pro 2018 with 2.2 GHz 6-Core Intel Core i7, and Acer Chromebook 11 with Intel Celeron N3060. For 720p input, the MacBook Pro can run the higher-quality model at 120 FPS and the end-to-end pipeline at 70 FPS, while the Chromebook runs inference at 62 FPS with the lower-quality model and 33 FPS end-to-end.
For quantitative evaluation of model accuracy, we adopt the popular metrics of intersection-over-union (IOU) and boundary F-measure. Both models achieve high quality, especially for having such a lightweight network:
We also release the accompanying Model Card for our segmentation models, which details our fairness evaluations. Our evaluation data contains images from 17 geographical subregions of the globe, with annotations for skin tone and gender. Our analysis shows that the model is consistent in its performance across the various regions, skin-tones, and genders, with only small deviations in IOU metrics.
Conclusion We introduced a new in-browser ML solution for blurring and replacing your background in Google Meet. With this, ML models and OpenGL shaders can run efficiently on the web. The developed features achieve real-time performance with low power consumption, even on low-power devices.
Acknowledgments Special thanks to those on the Meet team and others who worked on this project, in particular Sebastian Jansson, Rikard Lundmark, Stephan Reiter, Fabian Bergmark, Ben Wagner, Stefan Holmer, Dan Gunnarson, Stéphane Hulaud and to all our team members who worked on the technology with us: Siargey Pisarchyk, Karthik Raveendran, Chris McClanahan, Marat Dukhan, Frank Barchard, Ming Guang Yong, Chuo-Ling Chang, Michael Hays, Camillo Lugaresi, Gregory Karpiak, Siarhei Kazakou, Matsvei Zhdanovich, and Matthias Grundmann.
At Google, we're actively exploring how people can use creativity tools powered by machine learning and computational methods when producing multimedia content, from creating music and reframing videos, to drawing and more. One creative process in particular, video production, can especially benefit from such tools, as it requires a series of decisions about what content is best suited to a target audience, how to position the available assets within the field of view, and what temporal arrangement will yield the most compelling narrative. But what if one could leverage existing assets, such as a website, to get a jump-start on video creation? Businesses commonly host websites that contain rich visual representations about their services or products, all of which could be repurposed for other multimedia formats, such as videos, potentially enabling those without extensive resources the ability to reach a broader audience.
In “Automatic Video Creation From a Web Page”, published at UIST 2020, we introduce URL2Video, a research prototype pipeline to automatically convert a web page into a short video, given temporal and visual constraints provided by the content owner. URL2Video extracts assets (text, images, or videos) and their design styles (including fonts, colors, graphical layouts, and hierarchy) from HTML sources and organizes the visual assets into a sequence of shots, while maintaining a look-and-feel similar to the source page. Given a user-specified aspect ratio and duration, it then renders the repurposed materials into a video that is ideal for product and service advertising.
URL2Video Overview Assume a user provides an URL to a web page that illustrates their business. The URL2Video pipeline automatically selects key content from the page and decides the temporal and visual presentation of each asset, based on a set of heuristics derived from an interview study with designers who were familiar with web design and video ad creation. These designer-informed heuristics capture common video editing styles, including content hierarchy, constraining the amount of information in a shot and its time duration, providing consistent color and style for branding, and more. Using this information, the URL2Video pipeline parses a web page, analyzing the content and selecting visually salient text or images while preserving their design styles, which it organizes according to the video specifications provided by the user.
Webpage Analysis Given a webpage URL, URL2Video extracts document object model (DOM) information and multimedia materials. For the purposes of our research prototype, we limited the domain to static web pages that contain salient assets and headings preserved in an HTML hierarchy that follows recent web design principles, which encourage the use of prominent elements, distinct sections, and an order of visual focus that guides readers in perceiving information. URL2Video identifies such visually-distinguishable elements as a candidate list of asset groups, each of which may contain a heading, a product image, detailed descriptions, and call-to-action buttons, and captures both the raw assets (text and multimedia files) and detailed design specifications (HTML tags, CSS styles, and rendered locations) for each element. It then ranks the asset groups by assigning each a priority score based on their visual appearance and annotations, including their HTML tags, rendered sizes, and ordering shown on the page. In this way, an asset group that occupies a larger area at the top of the page receives a higher score.
Constraints-Based Asset Selection We consider two goals when composing a video: (1) each video shot should provide concise information, and (2) the visual design should be consistent with the source page. Based on these goals and the video constraints provided by the user, including the intended video duration (in seconds) and aspect ratio (commonly 16:9, 4:3, 1:1, etc.), URL2Video automatically selects and orders the asset groups to optimize the total priority score. To make the content concise, it presents only dominant elements from a page, such as a headline and a few multimedia assets. It constrains the duration of each visual element for viewers to perceive the content. In this way, a short video highlights the most salient information from the top of the page, and a longer video contains more campaigns or products.
Scene Composition & Video Rendering Given an ordered list of assets based on the DOM hierarchy, URL2Video follows the design heuristics obtained from interview studies to make decisions about both the temporal and spatial arrangement to present the assets in individual shots. It transfers the graphical layout of elements into the video’s aspect ratio, and applies the style choices including fonts and colors. To make a video more dynamic and engaging, it adjusts the presentation timing of assets. Finally, it renders the content into a video in the MPEG-4 container format.
User Control The interface to the research prototype allows the user to review the design attributes in each video shot extracted from the source page, reorder the materials, change the detailed design, such as colors and fonts, and adjust the constraints to generate a new video.
URL2Video Use Cases We demonstrate the performance of the end-to-end URL2Video pipeline on a variety of existing web pages. Below we highlight an example result where URL2Video converts a page that embeds multiple short video clips into a 12-second output video. Note how the pipeline makes automatic editing decisions on font and color choices, timing, and content ordering in a video captured from the source page.
The video below provides further demonstration:
To evaluate the automatically-generated videos, we conducted a user study with designers at Google. Our results show that URL2Video effectively extracted design elements from a web page and supported designers by bootstrapping the video creation process.
Next steps While this current research focuses on the visual presentation, we are developing new techniques that support the audio track and a voiceover in video editing. All in all, we envision a future where creators focus on making high-level decisions and an ML model interactively suggests detailed temporal and graphical edits for a final video creation on multiple platforms.
Acknowledgments We greatly thank our paper co-authors, Zheng Sun (Research) and Katrina Panovich (YouTube). We would also like to thank our colleagues who contributed to URL2Video, (in alphabetical order of last name) Jordan Canedy, Brian Curless, Nathan Frey, Madison Le, Alireza Mahdian, Justin Parra, Emily Ryan, Mogan Shieh, Sandor Szego, and Weilong Yang. We are grateful to receive the support from our leadership, Tomas Izo, Rahul Sukthankar, and Jay Yagnik.
