William Chan

Jonathan Ho, William Chan, Chitwan Saharia et al. · deepmind

We present Imagen Video, a text-conditional video generation system based on a cascade of video diffusion models. Given a text prompt, Imagen Video generates high definition videos using a base video generation model and a sequence of interleaved spatial and temporal video super-resolution models. We describe how we scale up the system as a high definition text-to-video model including design decisions such as the choice of fully-convolutional temporal and spatial super-resolution models at certain resolutions, and the choice of the v-parameterization of diffusion models. In addition, we confirm and transfer findings from previous work on diffusion-based image generation to the video generation setting. Finally, we apply progressive distillation to our video models with classifier-free guidance for fast, high quality sampling. We find Imagen Video not only capable of generating videos of high fidelity, but also having a high degree of controllability and world knowledge, including the ability to generate diverse videos and text animations in various artistic styles and with 3D object understanding. See https://imagen.research.google/video/ for samples.

Chitwan Saharia, William Chan, Saurabh Saxena et al. · deepmind

We present Imagen, a text-to-image diffusion model with an unprecedented degree of photorealism and a deep level of language understanding. Imagen builds on the power of large transformer language models in understanding text and hinges on the strength of diffusion models in high-fidelity image generation. Our key discovery is that generic large language models (e.g. T5), pretrained on text-only corpora, are surprisingly effective at encoding text for image synthesis: increasing the size of the language model in Imagen boosts both sample fidelity and image-text alignment much more than increasing the size of the image diffusion model. Imagen achieves a new state-of-the-art FID score of 7.27 on the COCO dataset, without ever training on COCO, and human raters find Imagen samples to be on par with the COCO data itself in image-text alignment. To assess text-to-image models in greater depth, we introduce DrawBench, a comprehensive and challenging benchmark for text-to-image models. With DrawBench, we compare Imagen with recent methods including VQ-GAN+CLIP, Latent Diffusion Models, and DALL-E 2, and find that human raters prefer Imagen over other models in side-by-side comparisons, both in terms of sample quality and image-text alignment. See https://imagen.research.google/ for an overview of the results.

Daniel Watson, William Chan, Ricardo Martin-Brualla et al. · deepmind

We present 3DiM, a diffusion model for 3D novel view synthesis, which is able to translate a single input view into consistent and sharp completions across many views. The core component of 3DiM is a pose-conditional image-to-image diffusion model, which takes a source view and its pose as inputs, and generates a novel view for a target pose as output. 3DiM can generate multiple views that are 3D consistent using a novel technique called stochastic conditioning. The output views are generated autoregressively, and during the generation of each novel view, one selects a random conditioning view from the set of available views at each denoising step. We demonstrate that stochastic conditioning significantly improves the 3D consistency of a naive sampler for an image-to-image diffusion model, which involves conditioning on a single fixed view. We compare 3DiM to prior work on the SRN ShapeNet dataset, demonstrating that 3DiM's generated completions from a single view achieve much higher fidelity, while being approximately 3D consistent. We also introduce a new evaluation methodology, 3D consistency scoring, to measure the 3D consistency of a generated object by training a neural field on the model's output views. 3DiM is geometry free, does not rely on hyper-networks or test-time optimization for novel view synthesis, and allows a single model to easily scale to a large number of scenes.

Rosanne Liu, Dan Garrette, Chitwan Saharia et al. · deepmind

Current image generation models struggle to reliably produce well-formed visual text. In this paper, we investigate a key contributing factor: popular text-to-image models lack character-level input features, making it much harder to predict a word's visual makeup as a series of glyphs. To quantify this effect, we conduct a series of experiments comparing character-aware vs. character-blind text encoders. In the text-only domain, we find that character-aware models provide large gains on a novel spelling task (WikiSpell). Applying our learnings to the visual domain, we train a suite of image generation models, and show that character-aware variants outperform their character-blind counterparts across a range of novel text rendering tasks (our DrawText benchmark). Our models set a much higher state-of-the-art on visual spelling, with 30+ point accuracy gains over competitors on rare words, despite training on far fewer examples.

Luyang Zhu, Dawei Yang, Tyler Zhu et al. · uw

Given two images depicting a person and a garment worn by another person, our goal is to generate a visualization of how the garment might look on the input person. A key challenge is to synthesize a photorealistic detail-preserving visualization of the garment, while warping the garment to accommodate a significant body pose and shape change across the subjects. Previous methods either focus on garment detail preservation without effective pose and shape variation, or allow try-on with the desired shape and pose but lack garment details. In this paper, we propose a diffusion-based architecture that unifies two UNets (referred to as Parallel-UNet), which allows us to preserve garment details and warp the garment for significant pose and body change in a single network. The key ideas behind Parallel-UNet include: 1) garment is warped implicitly via a cross attention mechanism, 2) garment warp and person blend happen as part of a unified process as opposed to a sequence of two separate tasks. Experimental results indicate that TryOnDiffusion achieves state-of-the-art performance both qualitatively and quantitatively.

Jonathan Ho, Tim Salimans, Alexey Gritsenko et al.

Generating temporally coherent high fidelity video is an important milestone in generative modeling research. We make progress towards this milestone by proposing a diffusion model for video generation that shows very promising initial results. Our model is a natural extension of the standard image diffusion architecture, and it enables jointly training from image and video data, which we find to reduce the variance of minibatch gradients and speed up optimization. To generate long and higher resolution videos we introduce a new conditional sampling technique for spatial and temporal video extension that performs better than previously proposed methods. We present the first results on a large text-conditioned video generation task, as well as state-of-the-art results on established benchmarks for video prediction and unconditional video generation. Supplementary material is available at https://video-diffusion.github.io/

Qingqing Huang, Daniel S. Park, Tao Wang et al.

We introduce Noise2Music, where a series of diffusion models is trained to generate high-quality 30-second music clips from text prompts. Two types of diffusion models, a generator model, which generates an intermediate representation conditioned on text, and a cascader model, which generates high-fidelity audio conditioned on the intermediate representation and possibly the text, are trained and utilized in succession to generate high-fidelity music. We explore two options for the intermediate representation, one using a spectrogram and the other using audio with lower fidelity. We find that the generated audio is not only able to faithfully reflect key elements of the text prompt such as genre, tempo, instruments, mood, and era, but goes beyond to ground fine-grained semantics of the prompt. Pretrained large language models play a key role in this story -- they are used to generate paired text for the audio of the training set and to extract embeddings of the text prompts ingested by the diffusion models. Generated examples: https://google-research.github.io/noise2music

Su Wang, Chitwan Saharia, Ceslee Montgomery et al.

Text-guided image editing can have a transformative impact in supporting creative applications. A key challenge is to generate edits that are faithful to input text prompts, while consistent with input images. We present Imagen Editor, a cascaded diffusion model built, by fine-tuning Imagen on text-guided image inpainting. Imagen Editor's edits are faithful to the text prompts, which is accomplished by using object detectors to propose inpainting masks during training. In addition, Imagen Editor captures fine details in the input image by conditioning the cascaded pipeline on the original high resolution image. To improve qualitative and quantitative evaluation, we introduce EditBench, a systematic benchmark for text-guided image inpainting. EditBench evaluates inpainting edits on natural and generated images exploring objects, attributes, and scenes. Through extensive human evaluation on EditBench, we find that object-masking during training leads to across-the-board improvements in text-image alignment -- such that Imagen Editor is preferred over DALL-E 2 and Stable Diffusion -- and, as a cohort, these models are better at object-rendering than text-rendering, and handle material/color/size attributes better than count/shape attributes.

Daniel Watson, William Chan, Jonathan Ho et al.

Diffusion models have emerged as an expressive family of generative models rivaling GANs in sample quality and autoregressive models in likelihood scores. Standard diffusion models typically require hundreds of forward passes through the model to generate a single high-fidelity sample. We introduce Differentiable Diffusion Sampler Search (DDSS): a method that optimizes fast samplers for any pre-trained diffusion model by differentiating through sample quality scores. We also present Generalized Gaussian Diffusion Models (GGDM), a family of flexible non-Markovian samplers for diffusion models. We show that optimizing the degrees of freedom of GGDM samplers by maximizing sample quality scores via gradient descent leads to improved sample quality. Our optimization procedure backpropagates through the sampling process using the reparametrization trick and gradient rematerialization. DDSS achieves strong results on unconditional image generation across various datasets (e.g., FID scores on LSUN church 128x128 of 11.6 with only 10 inference steps, and 4.82 with 20 steps, compared to 51.1 and 14.9 with strongest DDPM/DDIM baselines). Our method is compatible with any pre-trained diffusion model without fine-tuning or re-training required.

Chitwan Saharia, William Chan, Huiwen Chang et al.

This paper develops a unified framework for image-to-image translation based on conditional diffusion models and evaluates this framework on four challenging image-to-image translation tasks, namely colorization, inpainting, uncropping, and JPEG restoration. Our simple implementation of image-to-image diffusion models outperforms strong GAN and regression baselines on all tasks, without task-specific hyper-parameter tuning, architecture customization, or any auxiliary loss or sophisticated new techniques needed. We uncover the impact of an L2 vs. L1 loss in the denoising diffusion objective on sample diversity, and demonstrate the importance of self-attention in the neural architecture through empirical studies. Importantly, we advocate a unified evaluation protocol based on ImageNet, with human evaluation and sample quality scores (FID, Inception Score, Classification Accuracy of a pre-trained ResNet-50, and Perceptual Distance against original images). We expect this standardized evaluation protocol to play a role in advancing image-to-image translation research. Finally, we show that a generalist, multi-task diffusion model performs as well or better than task-specific specialist counterparts. Check out https://diffusion-palette.github.io for an overview of the results.

Yu Zhang, Daniel S. Park, Wei Han et al.

We summarize the results of a host of efforts using giant automatic speech recognition (ASR) models pre-trained using large, diverse unlabeled datasets containing approximately a million hours of audio. We find that the combination of pre-training, self-training and scaling up model size greatly increases data efficiency, even for extremely large tasks with tens of thousands of hours of labeled data. In particular, on an ASR task with 34k hours of labeled data, by fine-tuning an 8 billion parameter pre-trained Conformer model we can match state-of-the-art (SoTA) performance with only 3% of the training data and significantly improve SoTA with the full training set. We also report on the universal benefits gained from using big pre-trained and self-trained models for a large set of downstream tasks that cover a wide range of speech domains and span multiple orders of magnitudes of dataset sizes, including obtaining SoTA performance on many public benchmarks. In addition, we utilize the learned representation of pre-trained networks to achieve SoTA results on non-ASR tasks.

Nanxin Chen, Yu Zhang, Heiga Zen et al.

This paper introduces WaveGrad 2, a non-autoregressive generative model for text-to-speech synthesis. WaveGrad 2 is trained to estimate the gradient of the log conditional density of the waveform given a phoneme sequence. The model takes an input phoneme sequence, and through an iterative refinement process, generates an audio waveform. This contrasts to the original WaveGrad vocoder which conditions on mel-spectrogram features, generated by a separate model. The iterative refinement process starts from Gaussian noise, and through a series of refinement steps (e.g., 50 steps), progressively recovers the audio sequence. WaveGrad 2 offers a natural way to trade-off between inference speed and sample quality, through adjusting the number of refinement steps. Experiments show that the model can generate high fidelity audio, approaching the performance of a state-of-the-art neural TTS system. We also report various ablation studies over different model configurations. Audio samples are available at https://wavegrad.github.io/v2.

Daniel Watson, Jonathan Ho, Mohammad Norouzi et al.

Denoising Diffusion Probabilistic Models (DDPMs) have emerged as a powerful family of generative models that can yield high-fidelity samples and competitive log-likelihoods across a range of domains, including image and speech synthesis. Key advantages of DDPMs include ease of training, in contrast to generative adversarial networks, and speed of generation, in contrast to autoregressive models. However, DDPMs typically require hundreds-to-thousands of steps to generate a high fidelity sample, making them prohibitively expensive for high dimensional problems. Fortunately, DDPMs allow trading generation speed for sample quality through adjusting the number of refinement steps as a post process. Prior work has been successful in improving generation speed through handcrafting the time schedule by trial and error. We instead view the selection of the inference time schedules as an optimization problem, and introduce an exact dynamic programming algorithm that finds the optimal discrete time schedules for any pre-trained DDPM. Our method exploits the fact that ELBO can be decomposed into separate KL terms, and given any computation budget, discovers the time schedule that maximizes the training ELBO exactly. Our method is efficient, has no hyper-parameters of its own, and can be applied to any pre-trained DDPM with no retraining. We discover inference time schedules requiring as few as 32 refinement steps, while sacrificing less than 0.1 bits per dimension compared to the default 4,000 steps used on ImageNet 64x64 [Ho et al., 2020; Nichol and Dhariwal, 2021].

Jonathan Ho, Chitwan Saharia, William Chan et al.

We show that cascaded diffusion models are capable of generating high fidelity images on the class-conditional ImageNet generation benchmark, without any assistance from auxiliary image classifiers to boost sample quality. A cascaded diffusion model comprises a pipeline of multiple diffusion models that generate images of increasing resolution, beginning with a standard diffusion model at the lowest resolution, followed by one or more super-resolution diffusion models that successively upsample the image and add higher resolution details. We find that the sample quality of a cascading pipeline relies crucially on conditioning augmentation, our proposed method of data augmentation of the lower resolution conditioning inputs to the super-resolution models. Our experiments show that conditioning augmentation prevents compounding error during sampling in a cascaded model, helping us to train cascading pipelines achieving FID scores of 1.48 at 64x64, 3.52 at 128x128 and 4.88 at 256x256 resolutions, outperforming BigGAN-deep, and classification accuracy scores of 63.02% (top-1) and 84.06% (top-5) at 256x256, outperforming VQ-VAE-2.

Chitwan Saharia, Jonathan Ho, William Chan et al.

We present SR3, an approach to image Super-Resolution via Repeated Refinement. SR3 adapts denoising diffusion probabilistic models to conditional image generation and performs super-resolution through a stochastic denoising process. Inference starts with pure Gaussian noise and iteratively refines the noisy output using a U-Net model trained on denoising at various noise levels. SR3 exhibits strong performance on super-resolution tasks at different magnification factors, on faces and natural images. We conduct human evaluation on a standard 8X face super-resolution task on CelebA-HQ, comparing with SOTA GAN methods. SR3 achieves a fool rate close to 50%, suggesting photo-realistic outputs, while GANs do not exceed a fool rate of 34%. We further show the effectiveness of SR3 in cascaded image generation, where generative models are chained with super-resolution models, yielding a competitive FID score of 11.3 on ImageNet.

Edwin G. Ng, Chung-Cheng Chiu, Yu Zhang et al.

We combine recent advancements in end-to-end speech recognition to non-autoregressive automatic speech recognition. We push the limits of non-autoregressive state-of-the-art results for multiple datasets: LibriSpeech, Fisher+Switchboard and Wall Street Journal. Key to our recipe, we leverage CTC on giant Conformer neural network architectures with SpecAugment and wav2vec2 pre-training. We achieve 1.8%/3.6% WER on LibriSpeech test/test-other sets, 5.1%/9.8% WER on Switchboard, and 3.4% on the Wall Street Journal, all without a language model.

William Chan, Daniel Park, Chris Lee et al.

We present SpeechStew, a speech recognition model that is trained on a combination of various publicly available speech recognition datasets: AMI, Broadcast News, Common Voice, LibriSpeech, Switchboard/Fisher, Tedlium, and Wall Street Journal. SpeechStew simply mixes all of these datasets together, without any special re-weighting or re-balancing of the datasets. SpeechStew achieves SoTA or near SoTA results across a variety of tasks, without the use of an external language model. Our results include 9.0\% WER on AMI-IHM, 4.7\% WER on Switchboard, 8.3\% WER on CallHome, and 1.3\% on WSJ, which significantly outperforms prior work with strong external language models. We also demonstrate that SpeechStew learns powerful transfer learning representations. We fine-tune SpeechStew on a noisy low resource speech dataset, CHiME-6. We achieve 38.9\% WER without a language model, which compares to 38.6\% WER to a strong HMM baseline with a language model.

Harris Chan, Jamie Kiros, William Chan

A channel corresponds to a viewpoint or transformation of an underlying meaning. A pair of parallel sentences in English and French express the same underlying meaning, but through two separate channels corresponding to their languages. In this work, we present the Multichannel Generative Language Model (MGLM). MGLM is a generative joint distribution model over channels. MGLM marginalizes over all possible factorizations within and across all channels. MGLM endows flexible inference, including unconditional generation, conditional generation (where 1 channel is observed and other channels are generated), and partially observed generation (where incomplete observations are spread across all the channels). We experiment with the Multi30K dataset containing English, French, Czech, and German. We demonstrate experiments with unconditional, conditional, and partially conditional generation. We provide qualitative samples sampled unconditionally from the generative joint distribution. We also quantitatively analyze the quality-diversity trade-offs and find MGLM outperforms traditional bilingual discriminative models.

Nanxin Chen, Yu Zhang, Heiga Zen et al.

This paper introduces WaveGrad, a conditional model for waveform generation which estimates gradients of the data density. The model is built on prior work on score matching and diffusion probabilistic models. It starts from a Gaussian white noise signal and iteratively refines the signal via a gradient-based sampler conditioned on the mel-spectrogram. WaveGrad offers a natural way to trade inference speed for sample quality by adjusting the number of refinement steps, and bridges the gap between non-autoregressive and autoregressive models in terms of audio quality. We find that it can generate high fidelity audio samples using as few as six iterations. Experiments reveal WaveGrad to generate high fidelity audio, outperforming adversarial non-autoregressive baselines and matching a strong likelihood-based autoregressive baseline using fewer sequential operations. Audio samples are available at https://wavegrad.github.io/.

Chitwan Saharia, William Chan, Saurabh Saxena et al.

This paper presents two strong methods, CTC and Imputer, for non-autoregressive machine translation that model latent alignments with dynamic programming. We revisit CTC for machine translation and demonstrate that a simple CTC model can achieve state-of-the-art for single-step non-autoregressive machine translation, contrary to what prior work indicates. In addition, we adapt the Imputer model for non-autoregressive machine translation and demonstrate that Imputer with just 4 generation steps can match the performance of an autoregressive Transformer baseline. Our latent alignment models are simpler than many existing non-autoregressive translation baselines; for example, we do not require target length prediction or re-scoring with an autoregressive model. On the competitive WMT'14 En$\rightarrow$De task, our CTC model achieves 25.7 BLEU with a single generation step, while Imputer achieves 27.5 BLEU with 2 generation steps, and 28.0 BLEU with 4 generation steps. This compares favourably to the autoregressive Transformer baseline at 27.8 BLEU.

William Chan, Chitwan Saharia, Geoffrey Hinton et al.

This paper presents the Imputer, a neural sequence model that generates output sequences iteratively via imputations. The Imputer is an iterative generative model, requiring only a constant number of generation steps independent of the number of input or output tokens. The Imputer can be trained to approximately marginalize over all possible alignments between the input and output sequences, and all possible generation orders. We present a tractable dynamic programming training algorithm, which yields a lower bound on the log marginal likelihood. When applied to end-to-end speech recognition, the Imputer outperforms prior non-autoregressive models and achieves competitive results to autoregressive models. On LibriSpeech test-other, the Imputer achieves 11.1 WER, outperforming CTC at 13.0 WER and seq2seq at 12.5 WER.

Laura Ruis, Mitchell Stern, Julia Proskurnia et al.

We propose the Insertion-Deletion Transformer, a novel transformer-based neural architecture and training method for sequence generation. The model consists of two phases that are executed iteratively, 1) an insertion phase and 2) a deletion phase. The insertion phase parameterizes a distribution of insertions on the current output hypothesis, while the deletion phase parameterizes a distribution of deletions over the current output hypothesis. The training method is a principled and simple algorithm, where the deletion model obtains its signal directly on-policy from the insertion model output. We demonstrate the effectiveness of our Insertion-Deletion Transformer on synthetic translation tasks, obtaining significant BLEU score improvement over an insertion-only model.

Daniel S. Park, Yu Zhang, Chung-Cheng Chiu et al.

Recently, SpecAugment, an augmentation scheme for automatic speech recognition that acts directly on the spectrogram of input utterances, has shown to be highly effective in enhancing the performance of end-to-end networks on public datasets. In this paper, we demonstrate its effectiveness on tasks with large scale datasets by investigating its application to the Google Multidomain Dataset (Narayanan et al., 2018). We achieve improvement across all test domains by mixing raw training data augmented with SpecAugment and noise-perturbed training data when training the acoustic model. We also introduce a modification of SpecAugment that adapts the time mask size and/or multiplicity depending on the length of the utterance, which can potentially benefit large scale tasks. By using adaptive masking, we are able to further improve the performance of the Listen, Attend and Spell model on LibriSpeech to 2.2% WER on test-clean and 5.2% WER on test-other.

William Chan, Mitchell Stern, Jamie Kiros et al.

In this work, we present an empirical study of generation order for machine translation. Building on recent advances in insertion-based modeling, we first introduce a soft order-reward framework that enables us to train models to follow arbitrary oracle generation policies. We then make use of this framework to explore a large variety of generation orders, including uninformed orders, location-based orders, frequency-based orders, content-based orders, and model-based orders. Curiously, we find that for the WMT'14 English $\to$ German translation task, order does not have a substantial impact on output quality, with unintuitive orderings such as alphabetical and shortest-first matching the performance of a standard Transformer. This demonstrates that traditional left-to-right generation is not strictly necessary to achieve high performance. On the other hand, results on the WMT'18 English $\to$ Chinese task tend to vary more widely, suggesting that translation for less well-aligned language pairs may be more sensitive to generation order.

Lala Li, William Chan

The Insertion Transformer is well suited for long form text generation due to its parallel generation capabilities, requiring $O(\log_2 n)$ generation steps to generate $n$ tokens. However, modeling long sequences is difficult, as there is more ambiguity captured in the attention mechanism. This work proposes the Big Bidirectional Insertion Representations for Documents (Big BIRD), an insertion-based model for document-level translation tasks. We scale up the insertion-based models to long form documents. Our key contribution is introducing sentence alignment via sentence-positional embeddings between the source and target document. We show an improvement of +4.3 BLEU on the WMT'19 English$\rightarrow$German document-level translation task compared with the Insertion Transformer baseline.

Guan-Lin Chao, William Chan, Ian Lane

Speech recognition in cocktail-party environments remains a significant challenge for state-of-the-art speech recognition systems, as it is extremely difficult to extract an acoustic signal of an individual speaker from a background of overlapping speech with similar frequency and temporal characteristics. We propose the use of speaker-targeted acoustic and audio-visual models for this task. We complement the acoustic features in a hybrid DNN-HMM model with information of the target speaker's identity as well as visual features from the mouth region of the target speaker. Experimentation was performed using simulated cocktail-party data generated from the GRID audio-visual corpus by overlapping two speakers's speech on a single acoustic channel. Our audio-only baseline achieved a WER of 26.3%. The audio-visual model improved the WER to 4.4%. Introducing speaker identity information had an even more pronounced effect, improving the WER to 3.6%. Combining both approaches, however, did not significantly improve performance further. Our work demonstrates that speaker-targeted models can significantly improve the speech recognition in cocktail party environments.

William Chan, Nikita Kitaev, Kelvin Guu et al.

We present KERMIT, a simple insertion-based approach to generative modeling for sequences and sequence pairs. KERMIT models the joint distribution and its decompositions (i.e., marginals and conditionals) using a single neural network and, unlike much prior work, does not rely on a prespecified factorization of the data distribution. During training, one can feed KERMIT paired data $(x, y)$ to learn the joint distribution $p(x, y)$, and optionally mix in unpaired data $x$ or $y$ to refine the marginals $p(x)$ or $p(y)$. During inference, we have access to the conditionals $p(x \mid y)$ and $p(y \mid x)$ in both directions. We can also sample from the joint distribution or the marginals. The model supports both serial fully autoregressive decoding and parallel partially autoregressive decoding, with the latter exhibiting an empirically logarithmic runtime. We demonstrate through experiments in machine translation, representation learning, and zero-shot cloze question answering that our unified approach is capable of matching or exceeding the performance of dedicated state-of-the-art systems across a wide range of tasks without the need for problem-specific architectural adaptation.

Daniel S. Park, William Chan, Yu Zhang et al.

We present SpecAugment, a simple data augmentation method for speech recognition. SpecAugment is applied directly to the feature inputs of a neural network (i.e., filter bank coefficients). The augmentation policy consists of warping the features, masking blocks of frequency channels, and masking blocks of time steps. We apply SpecAugment on Listen, Attend and Spell networks for end-to-end speech recognition tasks. We achieve state-of-the-art performance on the LibriSpeech 960h and Swichboard 300h tasks, outperforming all prior work. On LibriSpeech, we achieve 6.8% WER on test-other without the use of a language model, and 5.8% WER with shallow fusion with a language model. This compares to the previous state-of-the-art hybrid system of 7.5% WER. For Switchboard, we achieve 7.2%/14.6% on the Switchboard/CallHome portion of the Hub5'00 test set without the use of a language model, and 6.8%/14.1% with shallow fusion, which compares to the previous state-of-the-art hybrid system at 8.3%/17.3% WER.

Jonathan Shen, Patrick Nguyen, Yonghui Wu et al.

Lingvo is a Tensorflow framework offering a complete solution for collaborative deep learning research, with a particular focus towards sequence-to-sequence models. Lingvo models are composed of modular building blocks that are flexible and easily extensible, and experiment configurations are centralized and highly customizable. Distributed training and quantized inference are supported directly within the framework, and it contains existing implementations of a large number of utilities, helper functions, and the newest research ideas. Lingvo has been used in collaboration by dozens of researchers in more than 20 papers over the last two years. This document outlines the underlying design of Lingvo and serves as an introduction to the various pieces of the framework, while also offering examples of advanced features that showcase the capabilities of the framework.

Mitchell Stern, William Chan, Jamie Kiros et al.

We present the Insertion Transformer, an iterative, partially autoregressive model for sequence generation based on insertion operations. Unlike typical autoregressive models which rely on a fixed, often left-to-right ordering of the output, our approach accommodates arbitrary orderings by allowing for tokens to be inserted anywhere in the sequence during decoding. This flexibility confers a number of advantages: for instance, not only can our model be trained to follow specific orderings such as left-to-right generation or a binary tree traversal, but it can also be trained to maximize entropy over all valid insertions for robustness. In addition, our model seamlessly accommodates both fully autoregressive generation (one insertion at a time) and partially autoregressive generation (simultaneous insertions at multiple locations). We validate our approach by analyzing its performance on the WMT 2014 English-German machine translation task under various settings for training and decoding. We find that the Insertion Transformer outperforms many prior non-autoregressive approaches to translation at comparable or better levels of parallelism, and successfully recovers the performance of the original Transformer while requiring only logarithmically many iterations during decoding.

Jianfeng Chi, Emmanuel Owusu, Xuwang Yin et al.

We present a practical method for protecting data during the inference phase of deep learning based on bipartite topology threat modeling and an interactive adversarial deep network construction. We term this approach \emph{Privacy Partitioning}. In the proposed framework, we split the machine learning models and deploy a few layers into users' local devices, and the rest of the layers into a remote server. We propose an approach to protect user's data during the inference phase, while still achieve good classification accuracy. We conduct an experimental evaluation of this approach on benchmark datasets of three computer vision tasks. The experimental results indicate that this approach can be used to significantly attenuate the capacity for an adversary with access to the state-of-the-art deep network's intermediate states to learn privacy-sensitive inputs to the network. For example, we demonstrate that our approach can prevent attackers from inferring the private attributes such as gender from the Face image dataset without sacrificing the classification accuracy of the original machine learning task such as Face Identification.

Bo Li, Yu Zhang, Tara Sainath et al.

We present two end-to-end models: Audio-to-Byte (A2B) and Byte-to-Audio (B2A), for multilingual speech recognition and synthesis. Prior work has predominantly used characters, sub-words or words as the unit of choice to model text. These units are difficult to scale to languages with large vocabularies, particularly in the case of multilingual processing. In this work, we model text via a sequence of Unicode bytes, specifically, the UTF-8 variable length byte sequence for each character. Bytes allow us to avoid large softmaxes in languages with large vocabularies, and share representations in multilingual models. We show that bytes are superior to grapheme characters over a wide variety of languages in monolingual end-to-end speech recognition. Additionally, our multilingual byte model outperform each respective single language baseline on average by 4.4% relatively. In Japanese-English code-switching speech, our multilingual byte model outperform our monolingual baseline by 38.6% relatively. Finally, we present an end-to-end multilingual speech synthesis model using byte representations which matches the performance of our monolingual baselines.

Sara Sabour, William Chan, Mohammad Norouzi

We present Optimal Completion Distillation (OCD), a training procedure for optimizing sequence to sequence models based on edit distance. OCD is efficient, has no hyper-parameters of its own, and does not require pretraining or joint optimization with conditional log-likelihood. Given a partial sequence generated by the model, we first identify the set of optimal suffixes that minimize the total edit distance, using an efficient dynamic programming algorithm. Then, for each position of the generated sequence, we use a target distribution that puts equal probability on the first token of all the optimal suffixes. OCD achieves the state-of-the-art performance on end-to-end speech recognition, on both Wall Street Journal and Librispeech datasets, achieving $9.3\%$ WER and $4.5\%$ WER respectively.

Takaaki Hori, Shinji Watanabe, Yu Zhang et al.

We present a state-of-the-art end-to-end Automatic Speech Recognition (ASR) model. We learn to listen and write characters with a joint Connectionist Temporal Classification (CTC) and attention-based encoder-decoder network. The encoder is a deep Convolutional Neural Network (CNN) based on the VGG network. The CTC network sits on top of the encoder and is jointly trained with the attention-based decoder. During the beam search process, we combine the CTC predictions, the attention-based decoder predictions and a separately trained LSTM language model. We achieve a 5-10\% error reduction compared to prior systems on spontaneous Japanese and Chinese speech, and our end-to-end model beats out traditional hybrid ASR systems.

William Chan, Yu Zhang, Quoc Le et al.

We present the Latent Sequence Decompositions (LSD) framework. LSD decomposes sequences with variable lengthed output units as a function of both the input sequence and the output sequence. We present a training algorithm which samples valid extensions and an approximate decoding algorithm. We experiment with the Wall Street Journal speech recognition task. Our LSD model achieves 12.9% WER compared to a character baseline of 14.8% WER. When combined with a convolutional network on the encoder, we achieve 9.6% WER.

Yu Zhang, William Chan, Navdeep Jaitly

Sequence-to-sequence models have shown success in end-to-end speech recognition. However these models have only used shallow acoustic encoder networks. In our work, we successively train very deep convolutional networks to add more expressive power and better generalization for end-to-end ASR models. We apply network-in-network principles, batch normalization, residual connections and convolutional LSTMs to build very deep recurrent and convolutional structures. Our models exploit the spectral structure in the feature space and add computational depth without overfitting issues. We experiment with the WSJ ASR task and achieve 10.5\% word error rate without any dictionary or language using a 15 layer deep network.

William Chan, Navdeep Jaitly, Quoc V. Le et al.

We present Listen, Attend and Spell (LAS), a neural network that learns to transcribe speech utterances to characters. Unlike traditional DNN-HMM models, this model learns all the components of a speech recognizer jointly. Our system has two components: a listener and a speller. The listener is a pyramidal recurrent network encoder that accepts filter bank spectra as inputs. The speller is an attention-based recurrent network decoder that emits characters as outputs. The network produces character sequences without making any independence assumptions between the characters. This is the key improvement of LAS over previous end-to-end CTC models. On a subset of the Google voice search task, LAS achieves a word error rate (WER) of 14.1% without a dictionary or a language model, and 10.3% with language model rescoring over the top 32 beams. By comparison, the state-of-the-art CLDNN-HMM model achieves a WER of 8.0%.

William Chan, Nan Rosemary Ke, Ian Lane

Deep Neural Network (DNN) acoustic models have yielded many state-of-the-art results in Automatic Speech Recognition (ASR) tasks. More recently, Recurrent Neural Network (RNN) models have been shown to outperform DNNs counterparts. However, state-of-the-art DNN and RNN models tend to be impractical to deploy on embedded systems with limited computational capacity. Traditionally, the approach for embedded platforms is to either train a small DNN directly, or to train a small DNN that learns the output distribution of a large DNN. In this paper, we utilize a state-of-the-art RNN to transfer knowledge to small DNN. We use the RNN model to generate soft alignments and minimize the Kullback-Leibler divergence against the small DNN. The small DNN trained on the soft RNN alignments achieved a 3.93 WER on the Wall Street Journal (WSJ) eval92 task compared to a baseline 4.54 WER or more than 13% relative improvement.

William Chan, Ian Lane

We present a novel deep Recurrent Neural Network (RNN) model for acoustic modelling in Automatic Speech Recognition (ASR). We term our contribution as a TC-DNN-BLSTM-DNN model, the model combines a Deep Neural Network (DNN) with Time Convolution (TC), followed by a Bidirectional Long Short-Term Memory (BLSTM), and a final DNN. The first DNN acts as a feature processor to our model, the BLSTM then generates a context from the sequence acoustic signal, and the final DNN takes the context and models the posterior probabilities of the acoustic states. We achieve a 3.47 WER on the Wall Street Journal (WSJ) eval92 task or more than 8% relative improvement over the baseline DNN models.

Haimonti Dutta, William Chan, Deepak Shankargouda et al.

The New York Public Library is participating in the Chronicling America initiative to develop an online searchable database of historically significant newspaper articles. Microfilm copies of the newspapers are scanned and high resolution Optical Character Recognition (OCR) software is run on them. The text from the OCR provides a wealth of data and opinion for researchers and historians. However, categorization of articles provided by the OCR engine is rudimentary and a large number of the articles are labeled editorial without further grouping. Manually sorting articles into fine-grained categories is time consuming if not impossible given the size of the corpus. This paper studies techniques for automatic categorization of newspaper articles so as to enhance search and retrieval on the archive. We explore unsupervised (e.g. KMeans) and semi-supervised (e.g. constrained clustering) learning algorithms to develop article categorization schemes geared towards the needs of end-users. A pilot study was designed to understand whether there was unanimous agreement amongst patrons regarding how articles can be categorized. It was found that the task was very subjective and consequently automated algorithms that could deal with subjective labels were used. While the small scale pilot study was extremely helpful in designing machine learning algorithms, a much larger system needs to be developed to collect annotations from users of the archive. The "BODHI" system currently being developed is a step in that direction, allowing users to correct wrongly scanned OCR and providing keywords and tags for newspaper articles used frequently. On successful implementation of the beta version of this system, we hope that it can be integrated with existing software being developed for the Chronicling America project.