In the era of large models, the autoregressive nature of decoding often results in latency serving as a significant bottleneck. We propose a non-autoregressive LM-fused ASR system that effectively leverages the parallelization capabilities of accelerator hardware. Our approach combines the Universal Speech Model (USM) and the PaLM 2 language model in per-segment scoring mode, achieving an average relative WER improvement across all languages of 10.8% on FLEURS and 3.6% on YouTube captioning. Furthermore, our comprehensive ablation study analyzes key parameters such as LLM size, context length, vocabulary size, fusion methodology. For instance, we explore the impact of LLM size ranging from 128M to 340B parameters on ASR performance. This study provides valuable insights into the factors influencing the effectiveness of practical large-scale LM-fused speech recognition systems.

Adapters are an efficient, composable alternative to full fine-tuning of pre-trained models and help scale the deployment of large ASR models to many tasks. In practice, a task ID is commonly prepended to the input during inference to route to single-task adapters for the specified task. However, one major limitation of this approach is that the task ID may not be known during inference, rendering it unsuitable for most multi-task settings. To address this, we propose three novel task-ID-free methods to combine single-task adapters in multi-task ASR and investigate two learning algorithms for training. We evaluate our methods on 10 test sets from 4 diverse ASR tasks and show that our methods are non-destructive and parameter-efficient. While only updating 17 % of the model parameters, our methods can achieve an 8 % mean WER improvement relative to full fine-tuning and are on-par with task-ID adapter routing.

In this work, we study the impact of Large-scale Language Models (LLM) on Automated Speech Recognition (ASR) of YouTube videos, which we use as a source for long-form ASR. We demonstrate up to 8% relative reduction in Word Error Eate (WER) on US English (en-us) and code-switched Indian English (en-in) long-form ASR test sets and a reduction of up to 30% relative on Salient Term Error Rate (STER) over a strong first-pass baseline that uses a maximum-entropy based language model. Improved lattice processing that results in a lattice with a proper (non-tree) digraph topology and carrying context from the 1-best hypothesis of the previous segment(s) results in significant wins in rescoring with LLMs. We also find that the gains in performance from the combination of LLMs trained on vast quantities of available data (such as C4 [1]) and conventional neural LMs is additive and significantly outperforms a strong first-pass baseline with a maximum entropy LM.

We explore unifying a neural segmenter with two-pass cascaded encoder ASR into a single model. A key challenge is allowing the segmenter (which runs in real-time, synchronously with the decoder) to finalize the non-causal 2nd pass (which runs 900 ms behind real-time) without introducing user-perceived latency or deletion errors during inference. We propose a design where the neural segmenter is integrated with the causal 1st pass decoder to emit a end-of-segment (EOS) signal in real-time. The EOS signal is then used to finalize the non-causal 2nd pass. We experiment with different ways to finalize the 2nd pass, and find that a dummy frame injection strategy allows for simultaneous high quality 2nd pass results and low finalization latency. On a real-world long-form captioning task (YouTube), we achieve 2.4% relative WER and 140 ms EOS latency gains over a baseline VAD-based segmenter with the same cascaded encoder.

Unpaired text and audio injection have emerged as dominant methods for improving ASR performance in the absence of a large labeled corpus. However, little guidance exists on deploying these methods to improve production ASR systems that are trained on very large supervised corpora and with realistic requirements like a constrained model size and CPU budget, streaming capability, and a rich lattice for rescoring and for downstream NLU tasks. In this work, we compare three state-of-the-art semi-supervised methods encompassing both unpaired text and audio as well as several of their combinations in a controlled setting using joint training. We find that in our setting these methods offer many improvements beyond raw WER, including substantial gains in tail-word WER, decoder computation during inference, and lattice density.

In this paper, we explore reducing computational latency of the 2-pass cascaded encoder model [1]. Specifically, we experiment with reducing the size of the causal 1st-pass and adding capacity to the non-causal 2nd-pass, such that the overall latency can be reduced without loss of quality. In addition, we explore using a confidence model for deciding to stop 2nd-pass recognition if we are confident in the 1st-pass hypothesis. Overall, we are able to reduce latency by a factor of 1.7X, compared to the baseline cascaded encoder from [1]. Secondly, with the added capacity in the non-causal 2nd-pass, we find that we can improve WER by up to 7% relative using wav2vec and minimum word-error-rate (MWER) training.

Capitalization normalization (truecasing) is the task of restoring the correct case (uppercase or lowercase) of noisy text. We propose a fast, accurate and compact two-level hierarchical word-and-character-based recurrent neural network model. We use the truecaser to normalize user-generated text in a Federated Learning framework for language modeling. A case-aware language model trained on this normalized text achieves the same perplexity as a model trained on text with gold capitalization. In a real user A/B experiment, we demonstrate that the improvement translates to reduced prediction error rates in a virtual keyboard application. Similarly, in an ASR language model fusion experiment, we show reduction in uppercase character error rate and word error rate.

Improving the performance of end-to-end ASR models on long utterances ranging from minutes to hours in length is an ongoing challenge in speech recognition. A common solution is to segment the audio in advance using a separate voice activity detector (VAD) that decides segment boundary locations based purely on acoustic speech/non-speech information. VAD segmenters, however, may be sub-optimal for real-world speech where, e.g., a complete sentence that should be taken as a whole may contain hesitations in the middle ("set an alarm for... 5 o'clock"). We propose to replace the VAD with an end-to-end ASR model capable of predicting segment boundaries in a streaming fashion, allowing the segmentation decision to be conditioned not only on better acoustic features but also on semantic features from the decoded text with negligible extra computation. In experiments on real world long-form audio (YouTube) with lengths of up to 30 minutes, we demonstrate 8.5% relative WER improvement and 250 ms reduction in median end-of-segment latency compared to the VAD segmenter baseline on a state-of-the-art Conformer RNN-T model.

End-to-end (E2E) models are often being accompanied by language models (LMs) via shallow fusion for boosting their overall quality as well as recognition of rare words.At the same time, several prior works show that LMs are susceptible to unintentionally memorizing rare or unique sequences in the training data.In this work, we design a framework for detecting memorization of random textual sequences (which we call canaries) in the LM training data when one has only black-box (query) access to LM-fused speech recognizer, as opposed to direct access to the LM.On a production-grade Conformer RNN-T E2E model fused with a Transformer LM, we show that detecting memorization of singly-occurring canaries from the LM training data of 300M examples is possible.Motivated to protect privacy, we also show that such memorization gets significantly reduced by per-example gradient-clipped LM training without compromising overall quality.

The careful construction of audio representations has become a dominant feature in the design of approaches to many speech tasks.Increasingly, such approaches have emphasized "disentanglement", where a representation contains only parts of the speech signal relevant to transcription while discarding irrelevant information.In this paper, we construct a representation learning task based on joint modeling of ASR and TTS, and seek to learn a representation of audio that disentangles that part of the speech signal that is relevant to transcription from that part which is not.We present empirical evidence that successfully finding such a representation is tied to the randomness inherent in training.We then make the observation that these desired, disentangled solutions to the optimization problem possess unique statistical properties.Finally, we show that enforcing these properties during training improves WER by 24.5% relative on average for our joint modeling task.These observations motivate a novel approach to learning effective audio representations.

Language models (LMs) significantly improve the recognition accuracy of end-to-end (E2E) models on words rarely seen during training, when used in either the shallow fusion or the rescoring setups. In this work, we introduce LMs in the learning of hybrid autoregressive transducer (HAT) models in the discriminative training framework, to mitigate the training versus inference gap regarding the use of LMs. For the shallow fusion setup, we use LMs during both hypotheses generation and loss computation, and the LM-aware MWER-trained model achieves 10\% relative improvement over the model trained with standard MWER on voice search test sets containing rare words. For the rescoring setup, we learn a small neural module to generate per-token fusion weights in a data-dependent manner. This model achieves the same rescoring WER as regular MWER-trained model, but without the need for sweeping fusion weights.

Text-only adaptation of a transducer model remains challenging for end-to-end speech recognition since the transducer has no clearly separated acoustic model (AM), language model (LM) or blank model. In this work, we propose a modular hybrid autoregressive transducer (MHAT) that has structurally separated label and blank decoders to predict label and blank distributions, respectively, along with a shared acoustic encoder. The encoder and label decoder outputs are directly projected to AM and internal LM scores and then added to compute label posteriors. We train MHAT with an internal LM loss and a HAT loss to ensure that its internal LM becomes a standalone neural LM that can be effectively adapted to text. Moreover, text adaptation of MHAT fosters a much better LM fusion than internal LM subtraction-based methods. On Google's large-scale production data, a multi-domain MHAT adapted with 100B sentences achieves relative WER reductions of up to 12.4% without LM fusion and 21.5% with LM fusion from 400K-hour trained HAT.

Dual learning is a paradigm for semi-supervised machine learning that seeks to leverage unsupervised data by solving two opposite tasks at once. In this scheme, each model is used to generate pseudo-labels for unlabeled examples that are used to train the other model. Dual learning has seen some use in speech processing by pairing ASR and TTS as dual tasks. However, these results mostly address only the case of using unpaired examples to compensate for very small supervised datasets, and mostly on large, non-streaming models. Dual learning has not yet been proven effective for using unsupervised data to improve realistic on-device streaming models that are already trained on large supervised corpora. We provide this missing piece though an analysis of an on-device-sized streaming conformer trained on the entirety of Librispeech, showing relative WER improvements of 10.7%/5.2% without an LM and 11.7%/16.4% with an LM.

Building ASR models across many languages is a challenging multi-task learning problem due to large variations and heavily unbalanced data. Existing work has shown positive transfer from high resource to low resource languages. However, degradations on high resource languages are commonly observed due to interference from the heterogeneous multilingual data and reduction in per-language capacity. We conduct a capacity study on a 15-language task, with the amount of data per language varying from 7.6K to 53.5K hours. We adopt GShard [1] to efficiently scale up to 10B parameters. Empirically, we find that (1) scaling the number of model parameters is an effective way to solve the capacity bottleneck - our 500M-param model already outperforms monolingual baselines and scaling it to 1B and 10B brought further quality gains; (2) larger models are not only more data efficient, but also more efficient in terms of training cost as measured in TPU days - the 1B-param model reaches the same accuracy at 34% of training time as the 500M-param model; (3) given a fixed capacity budget, adding depth works better than width and large encoders do better than large decoders; (4) with continuous training, they can be adapted to new languages and domains.

We introduce Lookup-Table Language Models (LookupLM), a method for scaling up the size of RNN language models with only a constant increase in the floating point operations, by increasing the expressivity of the embedding table. In particular, we instantiate an (additional) embedding table which embeds the previous n-gram token sequence, rather than a single token. This allows the embedding table to be scaled up arbitrarily -- with a commensurate increase in performance -- without changing the token vocabulary. Since embeddings are sparsely retrieved from the table via a lookup; increasing the size of the table adds neither extra operations to each forward pass nor extra parameters that need to be stored on limited GPU/TPU memory. We explore scaling n-gram embedding tables up to nearly a billion parameters. When trained on a 3-billion sentence corpus, we find that LookupLM improves long tail log perplexity by 2.44 and long tail WER by 23.4% on a downstream speech recognition task over a standard RNN language model baseline, an improvement comparable to a scaling up the baseline by 6.2x the number of floating point operations.

Innovations in neural architectures have fostered significant breakthroughs in language modeling and computer vision. Unfortunately, novel architectures often result in challenging hyper-parameter choices and training instability if the network parameters are not properly initialized. A number of architecture-specific initialization schemes have been proposed, but these schemes are not always portable to new architectures. This paper presents GradInit, an automated and architecture agnostic method for initializing neural networks. GradInit is based on a simple heuristic; the norm of each network layer is adjusted so that a single step of SGD or Adam with prescribed hyperparameters results in the smallest possible loss value. This adjustment is done by introducing a scalar multiplier variable in front of each parameter block, and then optimizing these variables using a simple numerical scheme. GradInit accelerates the convergence and test performance of many convolutional architectures, both with or without skip connections, and even without normalization layers. It also improves the stability of the original Transformer architecture for machine translation, enabling training it without learning rate warmup using either Adam or SGD under a wide range of learning rates and momentum coefficients. Code is available at https://github.com/zhuchen03/gradinit.

This paper studies how neural network architecture affects the speed of training. We introduce a simple concept called gradient confusion to help formally analyze this. When gradient confusion is high, stochastic gradients produced by different data samples may be negatively correlated, slowing down convergence. But when gradient confusion is low, data samples interact harmoniously, and training proceeds quickly. Through theoretical and experimental results, we demonstrate how the neural network architecture affects gradient confusion, and thus the efficiency of training. Our results show that, for popular initialization techniques, increasing the width of neural networks leads to lower gradient confusion, and thus faster model training. On the other hand, increasing the depth of neural networks has the opposite effect. Our results indicate that alternate initialization techniques or networks using both batch normalization and skip connections help reduce the training burden of very deep networks.

Data poisoning---the process by which an attacker takes control of a model by making imperceptible changes to a subset of the training data---is an emerging threat in the context of neural networks. Existing attacks for data poisoning neural networks have relied on hand-crafted heuristics, because solving the poisoning problem directly via bilevel optimization is generally thought of as intractable for deep models. We propose MetaPoison, a first-order method that approximates the bilevel problem via meta-learning and crafts poisons that fool neural networks. MetaPoison is effective: it outperforms previous clean-label poisoning methods by a large margin. MetaPoison is robust: poisoned data made for one model transfer to a variety of victim models with unknown training settings and architectures. MetaPoison is general-purpose, it works not only in fine-tuning scenarios, but also for end-to-end training from scratch, which till now hasn't been feasible for clean-label attacks with deep nets. MetaPoison can achieve arbitrary adversary goals---like using poisons of one class to make a target image don the label of another arbitrarily chosen class. Finally, MetaPoison works in the real-world. We demonstrate for the first time successful data poisoning of models trained on the black-box Google Cloud AutoML API.

Paper-intensive industries like insurance, law, and government have long leveraged optical character recognition (OCR) to automatically transcribe hordes of scanned documents into text strings for downstream processing. Even in 2019, there are still many scanned documents and mail that come into businesses in non-digital format. Text to be extracted from real world documents is often nestled inside rich formatting, such as tabular structures or forms with fill-in-the-blank boxes or underlines whose ink often touches or even strikes through the ink of the text itself. Further, the text region could have random ink smudges or spurious strokes. Such ink artifacts can severely interfere with the performance of recognition algorithms or other downstream processing tasks. In this work, we propose DeepErase, a neural-based preprocessor to erase ink artifacts from text images. We devise a method to programmatically assemble real text images and real artifacts into realistic-looking "dirty" text images, and use them to train an artifact segmentation network in a weakly supervised manner, since pixel-level annotations are automatically obtained during the assembly process. In addition to high segmentation accuracy, we show that our cleansed images achieve a significant boost in recognition accuracy by popular OCR software such as Tesseract 4.0. Finally, we test DeepErase on out-of-distribution datasets (NIST SDB) of scanned IRS tax return forms and achieve double-digit improvements in accuracy. All experiments are performed on both printed and handwritten text.

In this paper, we explore clean-label poisoning attacks on deep convolutional networks with access to neither the network’s output nor its architecture or parameters. Our goal is to ensure that after injecting the poisons into the training data, a model with unknown architecture and parameters trained on that data will misclassify the target image into a specific class. To achieve this goal, we generate multiple poison images from the base class by adding small perturbations which cause the poison images to trap the target image within their convex polytope in feature space. We also demonstrate that using Dropout during crafting of the poisons and enforcing this objective in multiple layers enhances transferability, enabling attacks against both the transfer learning and end-to-end training settings. We demonstrate transferable attack success rates of over 50% by poisoning only 1% of the training set.

Data poisoning is an attack on machine learning models wherein the attacker adds examples to the training set to manipulate the behavior of the model at test time. This paper explores poisoning attacks on neural nets. The proposed attacks use ``clean-labels''; they don't require the attacker to have any control over the labeling of training data. They are also targeted; they control the behavior of the classifier on a specific test instance without degrading overall classifier performance. For example, an attacker could add a seemingly innocuous image (that is properly labeled) to a training set for a face recognition engine, and control the identity of a chosen person at test time. Because the attacker does not need to control the labeling function, poisons could be entered into the training set simply by putting them online and waiting for them to be scraped by a data collection bot. We present an optimization-based method for crafting poisons, and show that just one single poison image can control classifier behavior when transfer learning is used. For full end-to-end training, we present a ``watermarking'' strategy that makes poisoning reliable using multiple (approx. 50) poisoned training instances. We demonstrate our method by generating poisoned frog images from the CIFAR dataset and using them to manipulate image classifiers.