Prabhat Pandey, Rupak Vignesh Swaminathan, K V Vijay Girish, Arunasish Sen, Jian. Xie, Grant Strimel, Andreas Schwarz. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Foundation models, such as large language models, have achieved remarkable success in natural language processing and are evolving into models capable of handling multiple modalities. Listening ability, in particular, is crucial for many applications, leading to research on building speech foundation models. However, the high computational cost of these large models presents a significant challenge for real-world applications. Although substantial efforts have been made to reduce computational costs, such as through pruning techniques, the majority of these approaches are applied primarily during the training phase for specific downstream tasks. In this study, we hypothesize that optimal pruned networks may vary based on contextual factors such as speaker characteristics, languages, and tasks. To address this, we propose a dynamic pruning technique that adapts to these contexts during inference without altering the underlying model. We demonstrated that we could successfully reduce inference time by approximately 30\% while maintaining accuracy in multilingual/multi-task scenarios. We also found that the obtained pruned structure offers meaningful interpretations based on the context, e.g., task-related information emerging as the dominant factor for efficient pruning.

The use of human speech to train LLMs poses privacy concerns due to these models' ability to generate samples that closely resemble artifacts in the training data. We propose a speaker privacy-preserving representation learning method through the Universal Speech Codec (USC), a computationally efficient codec that disentangles speech into: (i) privacy-preserving semantically rich representations, capturing content and speech paralinguistics, and (ii) residual acoustic and speaker representations that enable high-fidelity reconstruction. Evaluations show that USC's semantic representation preserves content, prosody, and sentiment, while removing identifiable traits. Additionally, we present an evaluation methodology for measuring privacy-preserving properties. We compare USC against other speech codecs and demonstrate its effectiveness on privacy-preserving representation learning, showcasing the trade-offs between speaker anonymization and paralinguistics retention.1

In this work, we propose a novel sequence-discriminative training criterion for automatic speech recognition (ASR) based on the Conformer Transducer. Inspired by the large-margin classifier framework, we separate the "good" and the "bad" hypotheses in an N-best list produced from a pre-trained transducer model by a margin (τ), hence the term, Max-Margin Transducer (MMT) loss. It is observed that fine-tuning with the proposed loss achieves significant improvement over baseline transducer loss but does not outperform the state-of-the-art minimum word error rate (MWER) training. However, combining the proposed MMT loss with MWER surpasses the performance of either losses suggesting the complimentary nature of MWER and MMT losses. With the combined losses, we obtained 7.44% and 7.68% relative WER improvements on Librispeech test-clean and test-other sets, respectively, and up to 8.9% relative improvement on Multi-lingual Librispeech test sets.

Neural contextual biasing for end-to-end neural ASR transducers has shown significant improvements in the recognition of named entities, such as contact names or device names. However, it comes with the cost of increased compute, as the biasing layers (which are usually based on cross-attention) add complexity to the neural transducers. In this paper, we propose gated contextual biasing models that can estimate at runtime when contextual biasing is needed and can toggle it on or off. That way, contextual biasing does not run on every audio frame, but only on the frames where it can be helpful for correct ASR recognition. We show that our gated contextual biasing models can maintain all the performance improvements of contextual biasing while offering significant compute-cost saving, as the contextual biasing needs to be executed for fewer than 15% of the audio frames.

Personalization in multi-turn dialogs has been a long standing challenge for end-to-end automatic speech recognition (E2E ASR) models. Recent work on contextual adapters has tackled rare word recognition using user catalogs. This adaptation, however, does not incorporate an important cue, the dialog act, which is available in a multi-turn dialog scenario. In this work, we propose a dialog act guided contextual adapter network. Specifically, it leverages dialog acts to select the most relevant user catalogs and creates queries based on both – the audio as well as the semantic relationship between the carrier phrase and user catalogs to better guide the contextual biasing. On industrial voice assistant datasets, our model outperforms both the baselines - dialog act encoder-only model, and the contextual adaptation, leading to the most improvement over the no-context model: 58% average relative word error rate reduction (WERR) in the multi-turn dialog scenario, in comparison to the prior-art contextual adapter, which has achieved 39% WERR over the no-context model.

Attention-based contextual biasing approaches have shown significant improvements in the recognition of generic and/or personal rare-words in End-to-End Automatic Speech Recognition (E2E ASR) systems like neural transducers. These approaches employ crossattention to bias the model towards specific contextual entities injected as bias-phrases to the model. Prior approaches typically relied on subword encoders for encoding the bias phrases. However, subword tokenizations are coarse and fail to capture granular pronunciation information which is crucial for biasing based on acoustic similarity. In this work, we propose to use lightweight character representations to encode fine-grained pronunciation features to improve contextual biasing guided by acoustic similarity between the audio and the contextual entities (termed acoustic biasing). We further integrate pretrained neural language model (NLM) based encoders to encode the utterance's semantic context along with contextual entities to perform biasing informed by the utterance’s semantic context (termed semantic biasing). Experiments using a Conformer Transducer model on the Librispeech dataset show a 4.62% - 9.26% relative WER improvement on different biasing list sizes over the baseline contextual model when incorporating our proposed acoustic and semantic biasing approach. On a large-scale in-house dataset, we observe 7.91% relative WER improvement compared to our baseline model. On tail utterances, the improvements are even more pronounced with 36.80% and 23.40% relative WER improvements on Librispeech rare words and an in-house testset respectively.

Tiny Signal-to-Interpretation (TinyS2I) has been recently introduced as an ultra low-footprint end-to-end spoken language understanding (SLU) model. This architecture is capable of running in ultra resource constrained environments like voice assistant devices, while at the same time reducing latency. In this work, we propose an extension to TinyS2I and train a multilingual system supporting several languages. Multilingual TinyS2I models show little to no degradation compared to their monolingual counterparts. Increasing the network size in width and depth improves the classification accuracy for mono- and multilingual setups, with the multilingual one improving beyond the monolingual accuracy. This enables users to interact with the device in the language of their choice and dynamically switch between languages without an explicit language setting or accuracy degradation.

End-to-End (E2E) automatic speech recognition (ASR) systems used in voice assistants often have difficulties recognizing infrequent words personalized to the user, such as names and places. Rare words often have non-trivial pronunciations, and in such cases, human knowledge in the form of a pronunciation lexicon can be useful. We propose a PROnunCiation-aware conTextual adaptER (PROCTER) that dynamically injects lexicon knowledge into an RNN-T model by adding a phonemic embedding along with a textual embedding. The experimental results show that the proposed PROCTER architecture outperforms the baseline RNN-T model by improving the word error rate (WER) by 44% and 57% when measured on personalized entities and personalized rare entities, respectively, while increasing the model size (number of trainable parameters) by only 1%. Furthermore, when evaluated in a zero-shot setting to recognize personalized device names, we observe 7% WER improvement with PROCTER, as compared to only 1% WER improvement with text-only contextual attention.

We present dual-attention neural biasing, an architecture designed to boost Wake Words (WW) recognition and improve inference time latency on speech recognition tasks. This architecture enables a dynamic switch for its runtime compute paths by exploiting WW spotting to select which branch of its attention networks to execute for an input audio frame. With this approach, we effectively improve WW spotting accuracy while saving runtime compute cost as defined by floating point operations (FLOPs). Using an in-house de-identified dataset, we demonstrate that the proposed dual-attention network can reduce the compute cost by 90% for WW audio frames, with only 1% increase in the number of parameters. This architecture improves WW F1 score by 16% relative and improves generic rare word error rate by 3% relative compared to the baselines.

Streaming speech recognition architectures are employed for low-latency, real-time applications. Such architectures are often characterized by their causality. Causal architectures emit tokens at each frame, relying only on current and past signal, while non-causal models are exposed to a window of future frames at each step to increase predictive accuracy. This dichotomy amounts to a trade-off for real-time Automatic Speech Recognition (ASR) system design: profit from the low-latency benefit of strictly-causal architectures while accepting predictive performance limitations, or realize the modeling benefits of future-context models accompanied by their higher latency penalty. In this work, we relax the constraints of this choice and present the Adaptive Non-Causal Attention Transducer (ANCAT). Our architecture is non-causal in the traditional sense, but executes in a low-latency, streaming manner by dynamically choosing when to rely on future context and to what degree within the audio stream. The resulting mechanism, when coupled with our novel regularization algorithms, delivers comparable accuracy to non-causal configurations while improving significantly upon latency, closing the gap with their causal counterparts. We showcase our design experimentally by reporting comparative ASR task results with measures of accuracy and latency on both publicly accessible and production-scale, voice-assistant datasets.

Recent years have seen significant advances in multi-turn Spoken Language Understanding (SLU), where dialogue contexts are used to guide intent classification and slot filling. However, how to selectively incorporate dialogue contexts, such as previous utterances and dialogue acts, into multi-turn SLU still remains a substantial challenge. In this work, we propose a novel contextual SLU model for multi-turn intent classification and slot filling tasks. We introduce an adaptive global-local context fusion mechanism to selectively integrate dialogue contexts into our model. The local context fusion aligns each dialogue context using multi-head attention, while the global context fusion measures overall context contribution to intent classification and slot filling tasks. Experiments show that on two benchmark datasets, our model achieves absolute F1 score improvements of 2.73% and 2.57% for the slot filling task on Sim-R and Sim M datasets, respectively. Additional experiments on a large-scale, de-identified, in-house dataset further verify the measurable accuracy gains of our proposed model.

We introduce Caching Networks (CachingNets), a speech recognition network architecture capable of delivering faster, more accurate decoding by leveraging common speech patterns. By explicitly incorporating select sentences unique to each user into the network’s design, we show how to train the model as an extension of the popular sequence transducer architecture through a multitask learning procedure. We further propose and experiment with different phrase caching policies, which are effective for virtual voice-assistant (VA) applications, to complement the architecture. Our results demonstrate that by pivoting between different inference strategies on the fly, CachingNets can deliver significant performance improvements. Specifically, on an industrial-scale, VA ASR task, we observe up to 7.4% relative word error rate (WER) and 11% sentence error rate (SER) improvements with accompanied latency gains.

On-device spoken language understanding (SLU) offers the potential for significant latency savings compared to cloud-based processing, as the audio stream does not need to be transmitted to a server. We present Tiny Signal-to-interpretation (TinyS2I), an end-to-end on-device SLU approach which is focused on heavily resource constrained devices. TinyS2I brings latency reduction without accuracy degradation, by exploiting use cases when the distribution of utterances that users speak to a device is largely heavy-tailed. The model is tailored to process on-device frequent utterances with support for dynamic contextual content, while deferring all other requests to the cloud. Compared to a powerful baseline, we demonstrate that TinyS2I achieves comparable performance, while offering latency gains due to local processing.

End-to-end Spoken Language Understanding (E2E SLU) has attracted increasing interest due to its advantages of joint optimization and low latency when compared to traditionally cascaded pipelines. Existing E2E SLU models usually follow a two-stage configuration where an Automatic Speech Recognition (ASR) network first predicts a transcript which is then passed to a Natural Language Understanding (NLU) module through an interface to infer semantic labels, such as intent and slot tags. This design, however, does not consider the NLU posterior while making transcript predictions, nor correct the NLU prediction error immediately by considering the previously predicted word-pieces. In addition, the NLU model in the two-stage system is not streamable, as it must wait for the audio segments to complete processing, which ultimately impacts the latency of the SLU system. In this work, we propose a streamable multi-task semantic transducer model to address these considerations. Our proposed architecture predicts ASR and NLU labels auto-regressively and uses a semantic decoder to ingest both previously predicted word-pieces and slot tags while aggregating them through a fusion network. Using an industry scale SLU and a public FSC dataset, we show the proposed model outperforms the two-stage E2E SLU model for both ASR and NLU metrics.

Personal rare word recognition in end-to-end Automatic Speech Recognition (E2E ASR) models is a challenge due to the lack of training data. A standard way to address this issue is with shallow fusion methods at inference time. However, due to their dependence on external language models and the deterministic approach to weight boosting, their performance is limited. In this paper, we propose training neural contextual adapters for personalization in neural transducer based ASR models. Our approach can not only bias towards user-defined words, but also has the flexibility to work with pretrained ASR models. Using an in-house dataset, we demonstrate that contextual adapters can be applied to any general purpose pretrained ASR model to improve personalization. Our method outperforms shallow fusion, while retaining functionality of the pretrained models by not altering any of the model weights. We further show that the adapter style training is superior to full-fine-tuning of the ASR models on datasets with user-defined content.

Dialogue act classification (DAC) is a critical task for spoken language understanding in dialogue systems. Prosodic features such as energy and pitch have been shown to be useful for DAC. Despite their importance, little research has explored neural approaches to integrate prosodic features into end-to-end (E2E) DAC models which infer dialogue acts directly from audio signals. In this work, we propose an E2E neural architecture that takes into account the need for characterizing prosodic phenomena co-occurring at different levels inside an utterance. A novel part of this architecture is a learnable gating mechanism that assesses the importance of prosodic features and selectively retains core information necessary for E2E DAC. Our proposed model improves DAC accuracy by 1.07% absolute across three publicly available benchmark datasets.

Conversational agents commonly utilize keyword spotting (KWS) to initiate voice interaction with the user. For user experience and privacy considerations, existing approaches to KWS largely focus on accuracy, which can often come at the expense of introduced latency. To address this tradeoff, we propose a novel approach to control KWS model latency and which generalizes to any loss function without explicit knowledge of the keyword endpoint. Through a single, tunable hyperparameter, our approach enables one to balance detection latency and accuracy for the targeted application. Empirically, we show that our approach gives superior performance under latency constraints when compared to existing methods. Namely, we make a substantial 25\% relative false accepts improvement for a fixed latency target when compared to the baseline state-of-the-art. We also show that when our approach is used in conjunction with a max-pooling loss, we are able to improve relative false accepts by 25 % at a fixed latency when compared to cross entropy loss.

We present a streaming, Transformer-based end-to-end automatic speech recognition (ASR) architecture which achieves efficient neural inference through compute cost amortization. Our architecture creates sparse computation pathways dynamically at inference time, resulting in selective use of compute resources throughout decoding, enabling significant reductions in compute with minimal impact on accuracy. The fully differentiable architecture is trained end-to-end with an accompanying lightweight arbitrator mechanism operating at the frame-level to make dynamic decisions on each input while a tunable loss function is used to regularize the overall level of compute against predictive performance. We report empirical results from experiments using the compute amortized Transformer-Transducer (T-T) model conducted on LibriSpeech data. Our best model can achieve a 60% compute cost reduction with only a 3% relative word error rate (WER) increase.

Machine learning model weights and activations are represented in full-precision during training. This leads to performance degradation in runtime when deployed on neural network accelerator (NNA) chips, which leverage highly parallelized fixed-point arithmetic to improve runtime memory and latency. In this work, we replicate the NNA operators during the training phase, accounting for the degradation due to low-precision inference on the NNA in back-propagation. Our proposed method efficiently emulates NNA operations, thus foregoing the need to transfer quantization error-prone data to the Central Processing Unit (CPU), ultimately reducing the user perceived latency (UPL). We apply our approach to Recurrent Neural Network-Transducer (RNN-T), an attractive architecture for on-device streaming speech recognition tasks. We train and evaluate models on 270K hours of English data and show a 5-7% improvement in engine latency while saving up to 10% relative degradation in WER.

For on-device automatic speech recognition (ASR), quantization aware training (QAT) is ubiquitous to achieve the trade-off between model predictive performance and efficiency. Among existing QAT methods, one major drawback is that the quantization centroids have to be predetermined and fixed. To overcome this limitation, we introduce a regularization-free, “soft-to-hard” compression mechanism with self-adjustable centroids in a$\mu$-Law constrained space, resulting in a simpler yet more versatile quantization scheme, called General Quantizer (GQ). We apply GQ to ASR tasks using Recurrent Neural Network Transducer (RNN-T) and Conformer architectures on both LibriSpeech and de-identified far-field datasets. Without accuracy degradation, GQ can compress both RNN-T and Conformer into sub-8-bit, and for some RNN-T layers, to 1-bit for fast and accurate inference. We observe a 30.73% memory footprint saving and 31.75% user-perceived latency reduction compared to 8-bit QAT via physical device benchmarking.

We present Bifocal RNN-T, a new variant of the Recurrent Neural Network Transducer (RNN-T) architecture designed for improved inference time latency on speech recognition tasks. The architecture enables a dynamic pivot for its runtime compute pathway, namely taking advantage of keyword spotting to select which component of the network to execute for a given audio frame. To accomplish this, we leverage a recurrent cell we call the Bifocal LSTM (BF-LSTM), which we detail in the paper. The architecture is compatible with other optimization strategies such as quantization, sparsification, and applying time-reduction layers, making it especially applicable for deployed, real-time speech recognition settings. We present the architecture and report comparative experimental results on voice-assistant speech recognition tasks. Specifically, we show our proposed Bifocal RNN-T can improve inference cost by 29.1% with matching word error rates and only a minor increase in memory size.

As more speech processing applications execute locally on edge devices, a set of resource constraints must be considered. In this work we address one of these constraints, namely over-the-network data budgets for transferring models from server to device. We present neural update approaches for release of subsequent speech model generations abiding by a data budget. We detail two architecture-agnostic methods which learn compact representations for transmission to devices. We experimentally validate our techniques with results on two tasks (automatic speech recognition and spoken language understanding) on open source data sets by demonstrating when applied in succession, our budgeted updates outperform comparable model compression baselines by significant margins.

We introduce Amortized Neural Networks (AmNets), a compute cost- and latency-aware network architecture particularly well-suited for sequence modeling tasks. We apply AmNets to the Recurrent Neural Network Transducer (RNN-T) to reduce compute cost and latency for an automatic speech recognition (ASR) task. The AmNets RNN-T architecture enables the network to dynamically switch between encoder branches on a frame-by-frame basis. Branches are constructed with variable levels of compute cost and model capacity. Here, we achieve variable compute for two well-known candidate techniques: one using sparse pruning and the other using matrix factorization. Frame-by-frame switching is determined by an arbitrator network that requires negligible compute overhead. We present results using both architectures on LibriSpeech data and show that our proposed architecture can reduce inference cost by up to 45\% and latency to nearly real-time without incurring a loss in accuracy.

We propose a simple yet effective method to compress an RNN-Transducer (RNN-T) through the well-known knowledge distillation paradigm. We show that the transducer's encoder outputs naturally have a high entropy and contain rich information about acoustically similar word-piece confusions. This rich information is suppressed when combined with the lower entropy decoder outputs to produce the joint network logits. Consequently, we introduce an auxiliary loss to distill the encoder logits from a teacher transducer's encoder, and explore training strategies where this encoder distillation works effectively. We find that tandem training of teacher and student encoders with an inplace encoder distillation outperforms the use of a pre-trained and static teacher transducer. We also report an interesting phenomenon we refer to as implicit distillation, that occurs when the teacher and student encoders share the same decoder. Our experiments show 5.37-8.4% relative word error rate reductions (WERR) on in-house test sets, and 5.05-6.18% relative WERRs on LibriSpeech test sets.

End-to-end spoken language understanding (SLU) models are a class of model architectures that predict semantics directly from speech. Because of their input and output types, we refer to them as speech-to-interpretation (STI) models. Previous works have successfully applied STI models to targeted use cases, such as recognizing home automation commands, however no study has yet addressed how these models generalize to broader use cases. In this work, we analyze the relationship between the performance of STI models and the difficulty of the use case to which they are applied. We introduce empirical measures of dataset semantic complexity to quantify the difficulty of the SLU tasks. We show that near-perfect performance metrics for STI models reported in the literature were obtained with datasets that have low semantic complexity values. We perform experiments where we vary the semantic complexity of a large, proprietary dataset and show that STI model performance correlates with our semantic complexity measures, such that performance increases as complexity values decrease. Our results show that it is important to contextualize an STI model's performance with the complexity values of its training dataset to reveal the scope of its applicability.

In this paper we investigate statistical model compression applied to natural language understanding (NLU) models.Smallfootprint NLU models are important for enabling offline systems on hardware restricted devices, and for decreasing ondemand model loading latency in cloud-based systems.To compress NLU models, we present two main techniques, parameter quantization and perfect feature hashing.These techniques are complementary to existing model pruning strategies such as L1 regularization.We performed experiments on a large scale NLU system.The results show that our approach achieves 14-fold reduction in memory usage compared to the original models with minimal predictive performance impact.

The robot coverage problem, a common planning problem, consists of finding a motion path for the robot that passes over all points in a given area or space. In many robotic applications involving coverage, e.g., industrial cleaning, mine sweeping, and agricultural operations, the desired coverage area is large and of arbitrary layout. In this work, we address the real problem of planning for coverage when the robot has limited battery or fuel, which restricts the length of travel of the robot before needing to be serviced. We introduce a new sweeping planning algorithm, which builds upon the boustrophedon cellular decomposition coverage algorithm to include a fixed fuel or battery capacity of the robot. We prove the algorithm is complete and show illustrative examples of the planned coverage outcome in a real building floor map.