Khe Chai Sim

Sandy Ritchie, You-Chi Cheng, Mingqing Chen et al.

Almost none of the 2,000+ languages spoken in Africa have widely available automatic speech recognition systems, and the required data is also only available for a few languages. We have experimented with two techniques which may provide pathways to large vocabulary speech recognition for African languages: multilingual modeling and self-supervised learning. We gathered available open source data and collected data for 15 languages, and trained experimental models using these techniques. Our results show that pooling the small amounts of data available in multilingual end-to-end models, and pre-training on unsupervised data can help improve speech recognition quality for many African languages.

Bo Li, Dongseong Hwang, Zhouyuan Huo et al.

Foundation models (FMs), that are trained on broad data at scale and are adaptable to a wide range of downstream tasks, have brought large interest in the research community. Benefiting from the diverse data sources such as different modalities, languages and application domains, foundation models have demonstrated strong generalization and knowledge transfer capabilities. In this paper, we present a pioneering study towards building an efficient solution for FM-based speech recognition systems. We adopt the recently developed self-supervised BEST-RQ for pretraining, and propose the joint finetuning with both source and unsupervised target domain data using JUST Hydra. The FM encoder adapter and decoder are then finetuned to the target domain with a small amount of supervised in-domain data. On a large-scale YouTube and Voice Search task, our method is shown to be both data and model parameter efficient. It achieves the same quality with only 21.6M supervised in-domain data and 130.8M finetuned parameters, compared to the 731.1M model trained from scratch on additional 300M supervised in-domain data.

Zhouyuan Huo, Khe Chai Sim, Bo Li et al.

Self-supervised pre-training of a speech foundation model, followed by supervised fine-tuning, has shown impressive quality improvements on automatic speech recognition (ASR) tasks. Fine-tuning separate foundation models for many downstream tasks are expensive since the foundation model is usually very big. Parameter-efficient fine-tuning methods (e.g. adapter, sparse update methods) offer an alternative paradigm where a small set of parameters are updated to adapt the foundation model to new tasks. However, these methods still suffer from a high computational memory cost and slow training speed because they require backpropagation through the entire neural network at each step. In the paper, we analyze the performance of features at different layers of a foundation model on the speech recognition task and propose a novel hierarchical feature fusion method for resource-efficient transfer learning from speech foundation models. Experimental results show that the proposed method can achieve better performance on speech recognition task than existing algorithms with fewer number of trainable parameters, less computational memory cost and faster training speed. After combining with Adapters at all layers, the proposed method can achieve the same performance as fine-tuning the whole model with $97\%$ fewer trainable encoder parameters and $53\%$ faster training speed.

Dongseong Hwang, Khe Chai Sim, Zhouyuan Huo et al.

State-of-the-art automatic speech recognition (ASR) systems are trained with tens of thousands of hours of labeled speech data. Human transcription is expensive and time consuming. Factors such as the quality and consistency of the transcription can greatly affect the performance of the ASR models trained with these data. In this paper, we show that we can train a strong teacher model to produce high quality pseudo labels by utilizing recent self-supervised and semi-supervised learning techniques. Specifically, we use JUST (Joint Unsupervised/Supervised Training) and iterative noisy student teacher training to train a 600 million parameter bi-directional teacher model. This model achieved 4.0% word error rate (WER) on a voice search task, 11.1% relatively better than a baseline. We further show that by using this strong teacher model to generate high-quality pseudo labels for training, we can achieve 13.6% relative WER reduction (5.9% to 5.1%) for a streaming model compared to using human labels.

Dongseong Hwang, Khe Chai Sim, Yu Zhang et al.

Knowledge distillation is an effective machine learning technique to transfer knowledge from a teacher model to a smaller student model, especially with unlabeled data. In this paper, we focus on knowledge distillation for the RNN-T model, which is widely used in state-of-the-art (SoTA) automatic speech recognition (ASR). Specifically, we compared using soft and hard target distillation to train large-scaleRNN-T models on the LibriSpeech/LibriLight public dataset (60k hours) and our in-house data (600k hours). We found that hard tar-gets are more effective when the teacher and student have different architecture, such as large teacher and small streaming student. On the other hand, soft target distillation works better in self-training scenario like iterative large teacher training. For a large model with0.6B weights, we achieve a new SoTA word error rate (WER) on LibriSpeech (8% relative improvement on dev-other) using Noisy Student Training with soft target distillation. It also allows our production teacher to adapt new data domain continuously.

Theresa Breiner, Swaroop Ramaswamy, Ehsan Variani et al.

Personalization of speech models on mobile devices (on-device personalization) is an active area of research, but more often than not, mobile devices have more text-only data than paired audio-text data. We explore training a personalized language model on text-only data, used during inference to improve speech recognition performance for that user. We experiment on a user-clustered LibriSpeech corpus, supplemented with personalized text-only data for each user from Project Gutenberg. We release this User-Specific LibriSpeech (UserLibri) dataset to aid future personalization research. LibriSpeech audio-transcript pairs are grouped into 55 users from the test-clean dataset and 52 users from test-other. We are able to lower the average word error rate per user across both sets in streaming and nonstreaming models, including an improvement of 2.5 for the harder set of test-other users when streaming.

Weiran Wang, Zelin Wu, Diamantino Caseiro et al.

Contextual biasing refers to the problem of biasing the automatic speech recognition (ASR) systems towards rare entities that are relevant to the specific user or application scenarios. We propose algorithms for contextual biasing based on the Knuth-Morris-Pratt algorithm for pattern matching. During beam search, we boost the score of a token extension if it extends matching into a set of biasing phrases. Our method simulates the classical approaches often implemented in the weighted finite state transducer (WFST) framework, but avoids the FST language altogether, with careful considerations on memory footprint and efficiency on tensor processing units (TPUs) by vectorization. Without introducing additional model parameters, our method achieves significant word error rate (WER) reductions on biasing test sets by itself, and yields further performance gain when combined with a model-based biasing method.

Liam Collins, Shanshan Wu, Sewoong Oh et al.

In many applications of federated learning (FL), clients desire models that are personalized using their local data, yet are also robust in the sense that they retain general global knowledge. However, the presence of data heterogeneity across clients induces a fundamental trade-off between personalization (i.e., adaptation to a local distribution) and robustness (i.e., not forgetting previously learned general knowledge). It is critical to understand how to navigate this personalization vs robustness trade-off when designing federated systems, which are increasingly moving towards a paradigm of fine-tuning large foundation models. Due to limited computational and communication capabilities in most federated settings, this foundation model fine-tuning must be done using parameter-efficient fine-tuning (PEFT) approaches. While some recent work has studied federated approaches to PEFT, the personalization vs robustness trade-off of federated PEFT has been largely unexplored. In this work, we take a step towards bridging this gap by benchmarking fundamental FL algorithms -- FedAvg and FedSGD plus personalization (via client local fine-tuning) -- applied to one of the most ubiquitous PEFT approaches to large language models (LLMs) -- prompt tuning -- in a multitude of hyperparameter settings under varying levels of data heterogeneity. Our results show that federated-trained prompts can be surprisingly robust when using a small learning rate with many local epochs for personalization, especially when using an adaptive optimizer as the client optimizer during federated training. We also demonstrate that simple approaches such as adding regularization and interpolating two prompts are effective in improving the personalization vs robustness trade-off in computation-limited settings with few local updates allowed for personalization.

Shefali Garg, Zhouyuan Huo, Khe Chai Sim et al.

Automatic speech recognition (ASR) systems have been shown to have large quality disparities between the language varieties they are intended or expected to recognize. One way to mitigate this is to train or fine-tune models with more representative datasets. But this approach can be hindered by limited in-domain data for training and evaluation. We propose a new way to improve the robustness of a US English short-form speech recognizer using a small amount of out-of-domain (long-form) African American English (AAE) data. We use CORAAL, YouTube and Mozilla Common Voice to train an audio classifier to approximately output whether an utterance is AAE or some other variety including Mainstream American English (MAE). By combining the classifier output with coarse geographic information, we can select a subset of utterances from a large corpus of untranscribed short-form queries for semi-supervised learning at scale. Fine-tuning on this data results in a 38.5% relative word error rate disparity reduction between AAE and MAE without reducing MAE quality.

Gemini Team, Petko Georgiev, Ving Ian Lei et al. · deepmind, mila

In this report, we introduce the Gemini 1.5 family of models, representing the next generation of highly compute-efficient multimodal models capable of recalling and reasoning over fine-grained information from millions of tokens of context, including multiple long documents and hours of video and audio. The family includes two new models: (1) an updated Gemini 1.5 Pro, which exceeds the February version on the great majority of capabilities and benchmarks; (2) Gemini 1.5 Flash, a more lightweight variant designed for efficiency with minimal regression in quality. Gemini 1.5 models achieve near-perfect recall on long-context retrieval tasks across modalities, improve the state-of-the-art in long-document QA, long-video QA and long-context ASR, and match or surpass Gemini 1.0 Ultra's state-of-the-art performance across a broad set of benchmarks. Studying the limits of Gemini 1.5's long-context ability, we find continued improvement in next-token prediction and near-perfect retrieval (>99%) up to at least 10M tokens, a generational leap over existing models such as Claude 3.0 (200k) and GPT-4 Turbo (128k). Finally, we highlight real-world use cases, such as Gemini 1.5 collaborating with professionals on completing their tasks achieving 26 to 75% time savings across 10 different job categories, as well as surprising new capabilities of large language models at the frontier; when given a grammar manual for Kalamang, a language with fewer than 200 speakers worldwide, the model learns to translate English to Kalamang at a similar level to a person who learned from the same content.

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Dongseong Hwang, Weiran Wang, Zhuoyuan Huo et al.

While Transformers have revolutionized deep learning, their quadratic attention complexity hinders their ability to process infinitely long inputs. We propose Feedback Attention Memory (FAM), a novel Transformer architecture that leverages a feedback loop to enable the network to attend to its own latent representations. This design fosters the emergence of working memory within the Transformer, allowing it to process indefinitely long sequences. TransformerFAM requires no additional weights, enabling seamless integration with pre-trained models. Our experiments show that TransformerFAM significantly improves Transformer performance on long-context tasks across various model sizes (1B, 8B, and 24B). These results showcase the potential to empower Large Language Models (LLMs) to process sequences of unlimited length.

Tsendsuren Munkhdalai, Youzheng Chen, Khe Chai Sim et al.

Parameter efficient adaptation methods have become a key mechanism to train large pre-trained models for downstream tasks. However, their per-task parameter overhead is considered still high when the number of downstream tasks to adapt for is large. We introduce an adapter module that has a better efficiency in large scale multi-task adaptation scenario. Our adapter is hierarchical in terms of how the adapter parameters are allocated. The adapter consists of a single shared controller network and multiple task-level adapter heads to reduce the per-task parameter overhead without performance regression on downstream tasks. The adapter is also recurrent so the entire adapter parameters are reused across different layers of the pre-trained model. Our Hierarchical Recurrent Adapter (HRA) outperforms the previous adapter-based approaches as well as full model fine-tuning baseline in both single and multi-task adaptation settings when evaluated on automatic speech recognition tasks.

Adam Stooke, Rohit Prabhavalkar, Khe Chai Sim et al.

Modern systems for automatic speech recognition, including the RNN-Transducer and Attention-based Encoder-Decoder (AED), are designed so that the encoder is not required to alter the time-position of information from the audio sequence into the embedding; alignment to the final text output is processed during decoding. We discover that the transformer-based encoder adopted in recent years is actually capable of performing the alignment internally during the forward pass, prior to decoding. This new phenomenon enables a simpler and more efficient model, the "Aligner-Encoder". To train it, we discard the dynamic programming of RNN-T in favor of the frame-wise cross-entropy loss of AED, while the decoder employs the lighter text-only recurrence of RNN-T without learned cross-attention -- it simply scans embedding frames in order from the beginning, producing one token each until predicting the end-of-message. We conduct experiments demonstrating performance remarkably close to the state of the art, including a special inference configuration enabling long-form recognition. In a representative comparison, we measure the total inference time for our model to be 2x faster than RNN-T and 16x faster than AED. Lastly, we find that the audio-text alignment is clearly visible in the self-attention weights of a certain layer, which could be said to perform "self-transduction".

Dongseong Hwang, Changwan Ryu, Khe Chai Sim

RNN-T is currently considered the industry standard in ASR due to its exceptional WERs in various benchmark tests and its ability to support seamless streaming and longform transcription. However, its biggest drawback lies in the significant discrepancy between its training and inference objectives. During training, RNN-T maximizes all alignment probabilities by teacher forcing, while during inference, it uses beam search which may not necessarily find the maximum probable alignment. Additionally, RNN-T's inability to experience mistakes during teacher forcing training makes it more problematic when a mistake occurs in inference. To address this issue, this paper proposes a Reinforcement Learning method that minimizes the gap between training and inference time. Our Edit Distance based RL (EDRL) approach computes rewards based on the edit distance, and trains the network at every action level. The proposed approach yielded SoTA WERs on LibriSpeech for the 600M Conformer RNN-T model.

Junwen Bai, Bo Li, Yu Zhang et al.

Self-supervised training has shown promising gains in pretraining models and facilitating the downstream finetuning for speech recognition, like multilingual ASR. Most existing methods adopt a 2-stage scheme where the self-supervised loss is optimized in the first pretraining stage, and the standard supervised finetuning resumes in the second stage. In this paper, we propose an end-to-end (E2E) Joint Unsupervised and Supervised Training (JUST) method to combine the supervised RNN-T loss and the self-supervised contrastive and masked language modeling (MLM) losses. We validate its performance on the public dataset Multilingual LibriSpeech (MLS), which includes 8 languages and is extremely imbalanced. On MLS, we explore (1) JUST trained from scratch, and (2) JUST finetuned from a pretrained checkpoint. Experiments show that JUST can consistently outperform other existing state-of-the-art methods, and beat the monolingual baseline by a significant margin, demonstrating JUST's capability of handling low-resource languages in multilingual ASR. Our average WER of all languages outperforms average monolingual baseline by 33.3%, and the state-of-the-art 2-stage XLSR by 32%. On low-resource languages like Polish, our WER is less than half of the monolingual baseline and even beats the supervised transfer learning method which uses external supervision.

Tsendsuren Munkhdalai, Khe Chai Sim, Angad Chandorkar et al.

Fast contextual adaptation has shown to be effective in improving Automatic Speech Recognition (ASR) of rare words and when combined with an on-device personalized training, it can yield an even better recognition result. However, the traditional re-scoring approaches based on an external language model is prone to diverge during the personalized training. In this work, we introduce a model-based end-to-end contextual adaptation approach that is decoder-agnostic and amenable to on-device personalization. Our on-device simulation experiments demonstrate that the proposed approach outperforms the traditional re-scoring technique by 12% relative WER and 15.7% entity mention specific F1-score in a continues personalization scenario.

Dongseong Hwang, Ananya Misra, Zhouyuan Huo et al.

Self- and semi-supervised learning methods have been actively investigated to reduce labeled training data or enhance the model performance. However, the approach mostly focus on in-domain performance for public datasets. In this study, we utilize the combination of self- and semi-supervised learning methods to solve unseen domain adaptation problem in a large-scale production setting for online ASR model. This approach demonstrates that using the source domain data with a small fraction of the target domain data (3%) can recover the performance gap compared to a full data baseline: relative 13.5% WER improvement for target domain data.

Zhouyuan Huo, Dongseong Hwang, Khe Chai Sim et al.

Streaming end-to-end speech recognition models have been widely applied to mobile devices and show significant improvement in efficiency. These models are typically trained on the server using transcribed speech data. However, the server data distribution can be very different from the data distribution on user devices, which could affect the model performance. There are two main challenges for on device training, limited reliable labels and limited training memory. While self-supervised learning algorithms can mitigate the mismatch between domains using unlabeled data, they are not applicable on mobile devices directly because of the memory constraint. In this paper, we propose an incremental layer-wise self-supervised learning algorithm for efficient speech domain adaptation on mobile devices, in which only one layer is updated at a time. Extensive experimental results demonstrate that the proposed algorithm obtains a Word Error Rate (WER) on the target domain $24.2\%$ better than supervised baseline and costs $89.7\%$ less training memory than the end-to-end self-supervised learning algorithm.

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.

Katrin Tomanek, Françoise Beaufays, Julie Cattiau et al.

While current state-of-the-art Automatic Speech Recognition (ASR) systems achieve high accuracy on typical speech, they suffer from significant performance degradation on disordered speech and other atypical speech patterns. Personalization of ASR models, a commonly applied solution to this problem, is usually performed in a server-based training environment posing problems around data privacy, delayed model-update times, and communication cost for copying data and models between mobile device and server infrastructure. In this paper, we present an approach to on-device based ASR personalization with very small amounts of speaker-specific data. We test our approach on a diverse set of 100 speakers with disordered speech and find median relative word error rate improvement of 71% with only 50 short utterances required per speaker. When tested on a voice-controlled home automation platform, on-device personalized models show a median task success rate of 81%, compared to only 40% of the unadapted models.

Mary Gooneratne, Khe Chai Sim, Petr Zadrazil et al.

Training machine learning models on mobile devices has the potential of improving both privacy and accuracy of the models. However, one of the major obstacles to achieving this goal is the memory limitation of mobile devices. Reducing training memory enables models with high-dimensional weight matrices, like automatic speech recognition (ASR) models, to be trained on-device. In this paper, we propose approximating the gradient matrices of deep neural networks using a low-rank parameterization as an avenue to save training memory. The low-rank gradient approximation enables more advanced, memory-intensive optimization techniques to be run on device. Our experimental results show that we can reduce the training memory by about 33.0% for Adam optimization. It uses comparable memory to momentum optimization and achieves a 4.5% relative lower word error rate on an ASR personalization task.

Khe Chai Sim, Françoise Beaufays, Arnaud Benard et al.

We study the effectiveness of several techniques to personalize end-to-end speech models and improve the recognition of proper names relevant to the user. These techniques differ in the amounts of user effort required to provide supervision, and are evaluated on how they impact speech recognition performance. We propose using keyword-dependent precision and recall metrics to measure vocabulary acquisition performance. We evaluate the algorithms on a dataset that we designed to contain names of persons that are difficult to recognize. Therefore, the baseline recall rate for proper names in this dataset is very low: 2.4%. A data synthesis approach we developed brings it to 48.6%, with no need for speech input from the user. With speech input, if the user corrects only the names, the name recall rate improves to 64.4%. If the user corrects all the recognition errors, we achieve the best recall of 73.5%. To eliminate the need to upload user data and store personalized models on a server, we focus on performing the entire personalization workflow on a mobile device.

Khe Chai Sim, Petr Zadrazil, Françoise Beaufays

Speaker-independent speech recognition systems trained with data from many users are generally robust against speaker variability and work well for a large population of speakers. However, these systems do not always generalize well for users with very different speech characteristics. This issue can be addressed by building personalized systems that are designed to work well for each specific user. In this paper, we investigate the idea of securely training personalized end-to-end speech recognition models on mobile devices so that user data and models never leave the device and are never stored on a server. We study how the mobile training environment impacts performance by simulating on-device data consumption. We conduct experiments using data collected from speech impaired users for personalization. Our results show that personalization achieved 63.7\% relative word error rate reduction when trained in a server environment and 58.1% in a mobile environment. Moving to on-device personalization resulted in 18.7% performance degradation, in exchange for improved scalability and data privacy. To train the model on device, we split the gradient computation into two and achieved 45% memory reduction at the expense of 42% increase in training time.

Yanzhang He, Tara N. Sainath, Rohit Prabhavalkar et al.

End-to-end (E2E) models, which directly predict output character sequences given input speech, are good candidates for on-device speech recognition. E2E models, however, present numerous challenges: In order to be truly useful, such models must decode speech utterances in a streaming fashion, in real time; they must be robust to the long tail of use cases; they must be able to leverage user-specific context (e.g., contact lists); and above all, they must be extremely accurate. In this work, we describe our efforts at building an E2E speech recognizer using a recurrent neural network transducer. In experimental evaluations, we find that the proposed approach can outperform a conventional CTC-based model in terms of both latency and accuracy in a number of evaluation categories.

Arun Narayanan, Ananya Misra, Khe Chai Sim et al.

Current state-of-the-art automatic speech recognition systems are trained to work in specific `domains', defined based on factors like application, sampling rate and codec. When such recognizers are used in conditions that do not match the training domain, performance significantly drops. This work explores the idea of building a single domain-invariant model for varied use-cases by combining large scale training data from multiple application domains. Our final system is trained using 162,000 hours of speech. Additionally, each utterance is artificially distorted during training to simulate effects like background noise, codec distortion, and sampling rates. Our results show that, even at such a scale, a model thus trained works almost as well as those fine-tuned to specific subsets: A single model can be robust to multiple application domains, and variations like codecs and noise. More importantly, such models generalize better to unseen conditions and allow for rapid adaptation -- we show that by using as little as 10 hours of data from a new domain, an adapted domain-invariant model can match performance of a domain-specific model trained from scratch using 70 times as much data. We also highlight some of the limitations of such models and areas that need addressing in future work.

Skanda Koppula, Khe Chai Sim, Kean Chin

We demonstrate a network visualization technique to analyze the recurrent state inside the LSTMs/GRUs used commonly in language and acoustic models. Interpreting intermediate state and network activations inside end-to-end models remains an open challenge. Our method allows users to understand exactly how much and what history is encoded inside recurrent state in grapheme sequence models. Our procedure trains multiple decoders that predict prior input history. Compiling results from these decoders, a user can obtain a signature of the recurrent kernel that characterizes its memory behavior. We demonstrate this method's usefulness in revealing information divergence in the bases of recurrent factorized kernels, visualizing the character-level differences between the memory of n-gram and recurrent language models, and extracting knowledge of history encoded in the layers of grapheme-based end-to-end ASR networks.

Bo Li, Tara N. Sainath, Khe Chai Sim et al.

Sequence-to-sequence models provide a simple and elegant solution for building speech recognition systems by folding separate components of a typical system, namely acoustic (AM), pronunciation (PM) and language (LM) models into a single neural network. In this work, we look at one such sequence-to-sequence model, namely listen, attend and spell (LAS), and explore the possibility of training a single model to serve different English dialects, which simplifies the process of training multi-dialect systems without the need for separate AM, PM and LMs for each dialect. We show that simply pooling the data from all dialects into one LAS model falls behind the performance of a model fine-tuned on each dialect. We then look at incorporating dialect-specific information into the model, both by modifying the training targets by inserting the dialect symbol at the end of the original grapheme sequence and also feeding a 1-hot representation of the dialect information into all layers of the model. Experimental results on seven English dialects show that our proposed system is effective in modeling dialect variations within a single LAS model, outperforming a LAS model trained individually on each of the seven dialects by 3.1 ~ 16.5% relative.