Large language models (LLM) have demonstrated the ability to understand human language by leveraging large amount of text data. Automatic speech recognition (ASR) systems are often limited by available transcribed speech data and benefit from a second pass rescoring using LLM. Recently multi-modal large language models, particularly speech and text foundational models have demonstrated strong spoken language understanding. Speech-Text foundational models leverage large amounts of unlabelled and labelled data both in speech and text modalities to model human language. In this work, we propose novel techniques to use multi-modal LLM for ASR rescoring. We also explore discriminative training to further improve the foundational model rescoring performance. We demonstrate cross-modal knowledge transfer in speech-text LLM can benefit rescoring. Our experiments demonstrate up-to 20% relative improvements over Whisper large ASR and up-to 15% relative improvements over text-only LLM.

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

While word error rates of automatic speech recognition (ASR) systems have consistently fallen, natural language understanding (NLU) applications built on top of ASR systems still attribute significant numbers of failures to low-quality speech recognition results. Existing assistant systems collect large numbers of these unsuccessful interactions, but these systems usually fail to learn from these interactions, even in an offline fashion. In this work, we introduce CLC: Contrastive Learning for Conversations, a family of methods for contrastive fine-tuning of models in a self-supervised fashion, making use of easily detectable artifacts in unsuccessful conversations with assistants. We demonstrate that our CLC family of approaches can improve the performance of ASR models on OD3, a new public large-scale semi-synthetic meta-dataset of audio task-oriented dialogues, by up to 19.2%. These gains transfer to real-world systems as well, where we show that CLC can help to improve performance by up to 6.7% over baselines.1

In the realm of spoken language understanding (SLU). numerous natural language understanding (NLU) methodologies have been adapted by supplying large language models (LLMs) with transcribed speech instead of conventional written text. In real-world scenarios, prior to input into an LLM. an automated speech recognition (ASR) system generates an output transcript hypothesis, where inherent errors can degrade subsequent SLU tasks. Here we introduce a method that utilizes the ASR system's lattice output instead of relying solely on the top hypothesis, aiming to encapsulate speech ambiguities and enhance SLU outcomes. Our in-context learning experiments, covering spoken question answering and intent classification. underline the LLM's resilience to noisy speech transcripts with the help of word confusion networks from lattices, bridging the SLU performance gap between using the top ASR hypothesis and an oracle upper bound. Additionally, we delve into the LLM's robustness to varying ASR performance conditions and scrutinize the aspects of in-context learning which prove the most influential.

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.

Dialog systems, such as voice assistants, are expected to engage with users in complex, evolving conversations. Unfortunately, traditional automatic speech recognition (ASR) systems deployed in such applications are usually trained to recognize each turn independently and lack the ability to adapt to the conversational context or incorporate user feedback. In this work, we introduce a general framework for ASR in dialog systems that can go beyond learning from single-turn utterances and learn over time how to adapt to both explicit supervision and implicit user feedback present in multi-turn conversations. We accomplish that by leveraging advances in student-teacher learning and context-aware dialog processing, and designing contrastive self-supervision approaches with Ohm, a new online hard-negative mining approach. We show that leveraging our new framework compared to traditional training leads to relative WER reductions of close to 10% in real-world dialog systems, and up to 26% on public synthetic data.

We introduce Condition-Aware Self-Supervised Learning Representation (CA-SSLR), a generalist conditioning model broadly applicable to various speech-processing tasks. Compared to standard fine-tuning methods that optimize for downstream models, CA-SSLR integrates language and speaker embeddings from earlier layers, making the SSL model aware of the current language and speaker context. This approach reduces the reliance on the input audio features while preserving the integrity of the base SSLR. CA-SSLR improves the model’s capabilities and demonstrates its generality on unseen tasks with minimal task-specific tuning. Our method employs linear modulation to dynamically adjust internal representations, enabling fine-grained adaptability without significantly altering the original model behavior. Experiments show that CA-SSLR reduces the number of trainable parameters, mitigates overfitting, and excels in under-resourced and unseen tasks. Specifically, CA-SSLR achieves a 10\% relative reduction in LID errors, a 37\% improvement in ASR CER on the ML-SUPERB benchmark, and a 27\% decrease in SV EER on VoxCeleb-1, demonstrating its effectiveness.

Endpoint (EP) detection is a key component of far-field speech recognition systems that assist the user through voice commands. The endpoint detector has to trade-off between accuracy and latency, since waiting longer reduces the cases of users being cut-off early. We propose a novel two-pass solution for endpointing, where the utterance endpoint detected from a first pass endpointer is verified by a 2nd-pass model termed EP Arbitrator. Our method improves the trade-off between early cut-offs and latency over a baseline endpointer, as tested on datasets including voice-assistant transactional queries, conversational speech, and the public SLURP corpus. We demonstrate that our method shows improvements regardless of the first-pass EP model used.

Second pass rescoring is a critical component of competitive automatic speech recognition (ASR) systems. Large language models have demonstrated their ability in using pre-trained information for better rescoring of ASR hypothesis. Discriminative training, directly optimizing the minimum word-error-rate (MWER) criterion typically improves rescoring. In this study, we propose and explore several discriminative fine-tuning schemes for pre-trained LMs. We propose two architectures based on different pooling strategies of output embeddings and compare with probability based MWER. We conduct detailed comparisons between pre-trained causal and bidirectional LMs in discriminative settings. Experiments on LibriSpeech demonstrate that all MWER training schemes are beneficial, giving additional gains upto 8.5% WER. Proposed pooling variants achieve lower latency while retaining most improvements. Finally, our study concludes that bidirectionality is better utilized with discriminative training.

We propose a neural language modeling system based on low-rank adaptation (LoRA) for speech recognition output rescoring. Although pretrained language models (LMs) like BERT have shown superior performance in second-pass rescoring, the high computational cost of scaling up the pretraining stage and adapting the pretrained models to specific domains limit their practical use in rescoring. Here we present a method based on low-rank decomposition to train a rescoring BERT model and adapt it to new domains using only a fraction (0.08%) of the pretrained parameters. These inserted matrices are optimized through a discriminative training objective along with a correlation-based regularization loss. The proposed low-rank adaptation RescoreBERT (LoRB) architecture is evaluated on LibriSpeech and internal datasets with decreased training times by factors between 5.4 and 3.6.

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.

Despite improvements to the generalization performance of automated speech recognition (ASR) models, specializing ASR models for downstream tasks remains a challenging task, primarily due to reduced data availability (necessitating increased data collection), and rapidly shifting data distributions (requiring more frequent model fine-tuning). In this work, we investigate the potential of leveraging external knowledge, particularly through off-policy generated text-to-speech key-value stores, to allow for flexible post-training adaptation to new data distributions. In our approach, audio embeddings captured from text-to-speech are used, along with semantic text embeddings, to bias ASR via an approximate k-nearest-neighbor (KNN) based attentive fusion step. Our experiments on LibiriSpeech and in-house voice assistant/search datasets show that the proposed approach can reduce domain adaptation time by up to 1K GPU-hours while providing up to 3% WER improvement compared to a fine-tuning baseline, suggesting a promising approach for adapting production ASR systems in challenging zero and few-shot scenarios.

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.

Automatic speech recognition (ASR) models with low-footprint are increasingly being deployed on edge devices for conversational agents, which enhances privacy. We study the problem of federated continual incremental learning for recurrent neural network-transducer (RNN-T) ASR models in the privacy-enhancing scheme of learning on-device, without access to ground truth human transcripts or machine transcriptions from a stronger ASR model. In particular, we study the performance of a self-learning based scheme, with a paired teacher model updated through an exponential moving average of ASR models. Further, we propose using possibly noisy weak-supervision signals such as feedback scores and natural language understanding semantics determined from user behavior across multiple turns in a session of interactions with the conversational agent. These signals are leveraged in a multitask policy-gradient training approach to improve the performance of self-learning for ASR. Finally, we show how catastrophic forgetting can be mitigated by combining on-device learning with a memory-replay approach using selected historical datasets. These innovations allow for 10% relative improvement in WER on new use cases with minimal degradation on other test sets in the absence of strong-supervision signals such as ground-truth transcriptions.

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.

In interactive automatic speech recognition (ASR) systems, low-latency requirements limit the amount of search space that can be explored during decoding, particularly in end-to-end neural ASR.In this paper, we present a novel streaming ASR architecture that outputs a confusion network while maintaining limited latency, as needed for interactive applications.We show that 1-best results of our model are on par with a comparable RNN-T system, while the richer hypothesis set allows secondpass rescoring to achieve 10-20% lower word error rate on the LibriSpeech task.We also show that our model outperforms a strong RNN-T baseline on a far-field voice assistant task.

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.

Second-pass rescoring is an important component in automatic speech recognition (ASR) systems that is used to improve the outputs from a first-pass decoder by implementing a lattice rescoring or n-best re-ranking. While pretraining with a masked language model (MLM) objective has received great success in various natural language understanding (NLU) tasks, it has not gained traction as a rescoring model for ASR. Specifically, training a bidirectional model like BERT on a discriminative objective such as minimum WER (MWER) has not been explored. Here we show how to train a BERT-based rescoring model with MWER loss, to incorporate the improvements of a discriminative loss into fine-tuning of deep bidirectional pretrained models for ASR. Specifically, we propose a fusion strategy that incorporates the MLM into the discriminative training process to effectively distill knowledge from a pretrained model. We further propose an alternative discriminative loss. This approach, which we call RescoreBERT, reduces WER by 6.6%/3.4% relative on the LibriSpeech clean/other test sets over a BERT baseline without discriminative objective. We also evaluate our method on an internal dataset from a conversational agent and find that it reduces both latency and WER (by 3 to 8% relative) over an LSTM rescoring model.

Spoken language understanding (SLU) systems extract both text transcripts and semantics associated with intents and slots from input speech utterances.SLU systems usually consist of (1) an automatic speech recognition (ASR) module, (2) an interface module that exposes relevant outputs from ASR, and (3) a natural language understanding (NLU) module.Interfaces in SLU systems carry information on text transcriptions or richer information like neural embeddings from ASR to NLU.In this paper, we study how interfaces affect joint-training for spoken language understanding.Most notably, we obtain the state-of-theart results on the publicly available 50-hr SLURP [1] dataset.We first leverage large-size pretrained ASR and NLU models that are connected by a text interface, and then jointly train both models via a sequence loss function.For scenarios where pretrained models are not utilized, the best results are obtained through a joint sequence loss training using richer neural interfaces.Finally, we show the overall diminishing impact of leveraging pretrained models with increased training data size.

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.

Incremental learning is one paradigm to enable model building and updating at scale with streaming data. For end-to-end automatic speech recognition (ASR) tasks, the absence of human annotated labels along with the need for privacy preserving policies for model building makes it a daunting challenge. Motivated by these challenges, in this paper we use a cloud based framework for production systems to demonstrate insights from privacy preserving incremental learning for automatic speech recognition (ILASR). By privacy preserving, we mean, usage of ephemeral data which are not human annotated. This system is a step forward for production level ASR models for incremental/continual learning that offers near real-time test-bed for experimentation in the cloud for end-to-end ASR, while adhering to privacy-preserving policies. We show that the proposed system can improve the production models significantly ($3%$) over a new time period of six months even in the absence of human annotated labels with varying levels of weak supervision and large batch sizes in incremental learning. This improvement is $20%$ over test sets with new words and phrases in the new time period. We demonstrate the effectiveness of model building in a privacy-preserving incremental fashion for ASR while further exploring the utility of having an effective teacher model and use of large batch sizes.

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.

End-to-end (E2E) automatic speech recognition (ASR) systems often have difficulty recognizing uncommon words, that appear infrequently in the training data. One promising method, to improve the recognition accuracy on such rare words, is to latch onto personalized/contextual information at inference. In this work, we present a novel context-aware transformer transducer (CATT) network that improves the state-of-the-art transformer-based ASR system by taking advantage of such contextual signals. Specifically, we propose a multi-head attention-based context-biasing network, which is jointly trained with the rest of the ASR sub-networks. We explore different techniques to encode contextual data and to create the final attention context vectors. We also leverage both BLSTM and pretrained BERT based models to encode contextual data and guide the network training. Using an in-house far-field dataset, we show that CATT, using a BERT based context encoder, improves the word error rate of the baseline transformer transducer and outperforms an existing deep contextual model by 24.2% and 19.4% respectively.

The recognition of personalized content, such as contact names, remains a challenging problem for end-to-end speech recognition systems. In this work, we demonstrate how first- and second-pass rescoring strategies can be leveraged together to improve the recognition of such words. Following previous work, we use a shallow fusion approach to bias towards recognition of personalized content in the first-pass decoding. We show that such an approach can improve personalized content recognition by up to 16% with minimum degradation on the general use case. We describe a fast and scalable algorithm that enables our biasing models to remain at the word-level, while applying the biasing at the subword level. This has the advantage of not requiring the biasing models to be dependent on any subword symbol table. We also describe a novel second-pass de-biasing approach: used in conjunction with a first-pass shallow fusion that optimizes on oracle WER, we can achieve an additional 14% improvement on personalized content recognition, and even improve accuracy for the general use case by up to 2.5%.

Accents mismatching is a critical problem for end-to-end ASR. This paper aims to address this problem by building an accent-robust RNN-T system with domain adversarial training (DAT). We unveil the magic behind DAT and provide, for the first time, a theoretical guarantee that DAT learns accent-invariant representations. We also prove that performing the gradient reversal in DAT is equivalent to minimizing the Jensen-Shannon divergence between domain output distributions. Motivated by the proof of equivalence, we introduce reDAT, a novel technique based on DAT, which relabels data using either unsupervised clustering or soft labels. Experiments on 23K hours of multi-accent data show that DAT achieves competitive results over accent-specific baselines on both native and non-native English accents but up to 13% relative WER reduction on unseen accents; our reDAT yields further improvements over DAT by 3% and 8% relatively on non-native accents of American and British English.

As voice assistants become more ubiquitous, they are increasingly expected to support and perform well on a wide variety of use-cases across different domains. We present a domain-aware rescoring framework suitable for achieving domain-adaptation during second-pass rescoring in production settings. In our framework, we fine-tune a domain-general neural language model on several domains, and use an LSTM-based domain classification model to select the appropriate domain-adapted model to use for second-pass rescoring. This domain-aware rescoring improves the word error rate by up to 2.4% and slot word error rate by up to 4.1% on three individual domains – shopping, navigation, and music – compared to domain general rescoring. These improvements are obtained while maintaining accuracy for the general use case.

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.

Conventional dynamic language switching enables seamless multilingual interactions by running several monolingual ASR systems in parallel and triggering the appropriate downstream components using a standalone language identification (LID) service. Since this solution is neither scalable nor cost- and memory-efficient, especially for on-device applications, we propose end-to-end, streaming, joint ASR-LID architectures based on the recurrent neural network transducer framework. Two key formulations are explored: (1) joint training using a unified output space for ASR and LID vocabularies, and (2) joint training viewed as multi-task optimization. We also evaluate the benefit of using auxiliary language information obtained on-the-fly from an acoustic LID classifier. Experiments with the English-Hindi language pair show that: (a) multi-task architectures perform better overall, and (b) the best joint architecture surpasses monolingual ASR (6.4–9.2% word error rate reduction) and acoustic LID (53.9–56.1% error rate reduction) baselines while reducing the overall memory footprint by up to 46%.

Spoken language understanding (SLU) systems extract transcriptions, as well as semantics of intent or named entities from speech, and are essential components of voice activated systems. SLU models, which either directly extract semantics from audio or are composed of pipelined automatic speech recognition (ASR) and natural language understanding (NLU) models, are typically trained via differentiable cross-entropy losses, even when the relevant performance metrics of interest are word or semantic error rates. In this work, we propose non-differentiable sequence losses based on SLU metrics as a proxy for semantic error and use the REINFORCE trick to train ASR and SLU models with this loss. We show that custom sequence loss training is the state-of-the-art on open SLU datasets and leads to 6% relative improvement in both ASR and NLU performance metrics on large proprietary datasets. We also demonstrate how the semantic sequence loss training paradigm can be used to update ASR and SLU models without transcripts, using semantic feedback alone.

In order to achieve high accuracy for machine learning (ML) applications, it is essential to employ models with a large number of parameters. Certain applications, such as Automatic Speech Recognition (ASR), however, require real-time interactions with users, hence compelling the model to have as low latency as possible. Deploying large scale ML applications thus necessitates model quantization and compression, especially when running ML models on resource constrained devices. For example, by forcing some of the model weight values into zero, it is possible to apply zero-weight compression, which reduces both the model size and model reading time from the memory. In the literature, such methods are referred to as sparse pruning. The fundamental questions are when and which weights should be forced to zero, i.e. be pruned. In this work, we propose a compressed sensing based pruning (CSP) approach to effectively address those questions. By reformulating sparse pruning as a sparsity inducing and compression-error reduction dual problem, we introduce the classic compressed sensing process into the ML model training process. Using ASR task as an example, we show that CSP consistently outperforms existing approaches in the literature.

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.

Spoken language understanding (SLU) systems translate voice input commands to semantics which are encoded as an intent and pairs of slot tags and values.Most current SLU systems deploy a cascade of two neural models where the first one maps the input audio to a transcript (ASR) and the second predicts the intent and slots from the transcript (NLU).In this paper, we introduce FANS, a new end-to-end SLU model that fuses an ASR audio encoder to a multi-task NLU decoder to infer the intent, slot tags, and slot values directly from a given input audio, obviating the need for transcription.FANS consists of a shared audio encoder and three decoders, two of which are seq-to-seq decoders that predict non null slot tags and slot values in parallel and in an auto-regressive manner.FANS neural encoder and decoders architectures are flexible which allows us to leverage different combinations of LSTM, self-attention, and attenders.Our experiments show compared to the state-of-the-art end-toend SLU models, FANS reduces ICER and IRER errors relatively by 30% and 7%, respectively, when tested on an in-house SLU dataset and by 0.86% and 2% absolute when tested on a public SLU dataset.

Comprehending the overall intent of an utterance helps a listener recognize the individual words spoken. Inspired by this fact, we perform a novel study of the impact of explicitly incorporating intent representations as additional information to improve a recurrent neural network-transducer (RNN-T) based automatic speech recognition (ASR) system. An audio-to-intent (A2I) model encodes the intent of the utterance in the form of embeddings or posteriors, and these are used as auxiliary inputs for RNN-T training and inference. Experimenting with a 50k-hour far-field English speech corpus, this study shows that when running the system in non-streaming mode, where intent representation is extracted from the entire utterance and then used to bias streaming RNN-T search from the start, it provides a 5.56% relative word error rate reduction (WERR). On the other hand, a streaming system using per-frame intent posteriors as extra inputs for the RNN-T ASR system yields a 3.33% relative WERR. A further detailed analysis of the streaming system indicates that our proposed method brings especially good gain on media-playing related intents (e.g. 9.12% relative WERR on PlayMusicIntent).

Wav2vec-C introduces a novel representation learning technique combining elements from wav2vec 2.0 and VQ-VAE. Our model learns to reproduce quantized representations from partially masked speech encoding using a contrastive loss in a way similar to Wav2vec 2.0. However, the quantization process is regularized by an additional consistency network that learns to reconstruct the input features to the wav2vec 2.0 network from the quantized representations in a way similar to a VQ-VAE model. The proposed self-supervised model is trained on 10k hours of unlabeled data and subsequently used as the speech encoder in a RNN-T ASR model and fine-tuned with 1k hours of labeled data. This work is one of only a few studies of self-supervised learning on speech tasks with a large volume of real far-field labeled data. The Wav2vec-C encoded representations achieves, on average, twice the error reduction over baseline and a higher codebook utilization in comparison to wav2vec 2.0

In this paper, we propose a streaming model to distinguish voice queries intended for a smart-home device from background speech. The proposed model consists of multiple CNN layers with residual connections, followed by a stacked LSTM architecture. The streaming capability is achieved by using unidirectional LSTM layers and a causal mean aggregation layer to form the final utterance-level prediction up to the current frame. In order to avoid redundant computation during online streaming inference, we use a caching mechanism for every convolution operation. Experimental results on a device-directed vs. non device-directed task show that the proposed model yields an equal error rate reduction of 41% compared to our previous best model on this task. Furthermore, we show that the proposed model is able to accurately predict earlier in time compared to the attention-based models.

End-to-end approaches for automatic speech recognition (ASR) benefit from directly modeling the probability of the word sequence given the input audio stream in a single neural network. However, compared to conventional ASR systems, these models typically require more data to achieve comparable results. Well-known model adaptation techniques, to account for domain and style adaptation, are not easily applicable to end-to-end systems. Conventional HMM-based systems, on the other hand, have been optimized for various production environments and use cases. In this work, we propose to combine the benefits of end-to-end approaches with a conventional system using an attention-based discriminative language model that learns to rescore the output of a first-pass ASR system. We show that learning to rescore a list of potential ASR outputs is much simpler than learning to generate the hypothesis. The proposed model results in up to 8% improvement in word error rate even when the amount of training data is a fraction of data used for training the first-pass system.

We consider the problem of spoken language understanding (SLU) of extracting natural language intents and associated slot arguments or named entities from speech that is primarily directed at voice assistants. Such a system subsumes both automatic speech recognition (ASR) as well as natural language understanding (NLU). An end-to-end joint SLU model can be built to a required specification opening up the opportunity to deploy on hardware constrained scenarios like devices enabling voice assistants to work offline, in a privacy preserving manner, whilst also reducing server costs. We first present models that extract utterance intent directly from speech without intermediate text output. We then present a compositional model, which generates the transcript using the Listen Attend Spell ASR system and then extracts interpretation using a neural NLU model. Finally, we contrast these methods to a jointly trained end-to-end joint SLU model, consisting of ASR and NLU subsystems which are connected by a neural network based interface instead of text, that produces transcripts as well as NLU interpretation. We show that the jointly trained model shows improvements to ASR incorporating semantic information from NLU and also improves NLU by exposing it to ASR confusion encoded in the hidden layer.

Neural language models (NLM) have been shown to outperform conventional n-gram language models by a substantial margin in Automatic Speech Recognition (ASR) and other tasks. There are, however, a number of challenges that need to be addressed for an NLM to be used in a practical large-scale ASR system. In this paper, we present solutions to some of the challenges, including training NLM from heterogenous corpora, limiting latency impact and handling personalized bias in the second-pass rescorer. Overall, we show that we can achieve a 6.2% relative WER reduction using neural LM in a second-pass n-best rescoring framework with a minimal increase in latency.

Grapheme-to-phoneme (G2P) models are a key component in Automatic Speech Recognition (ASR) systems, such as the ASR system in Alexa, as they are used to generate pronunciations for out-of-vocabulary words that do not exist in the pronunciation lexicons (mappings like e c h o to E k oU). Most G2P systems are monolingual and based on traditional joint-sequence based n-gram models [1,2]. As an alternative, we present a single end-to-end trained neural G2P model that shares same encoder and decoder across multiple languages. This allows the model to utilize a combination of universal symbol inventories of Latin-like alphabets and cross-linguistically shared feature representations. Such model is especially useful in the scenarios of low resource languages and code switching/foreign words, where the pronunciations in one language need to be adapted to other locales or accents. We further experiment with word language distribution vector as an additional training target in order to improve system performance by helping the model decouple pronunciations across a variety of languages in the parameter space. We show 7.2% average improvement in phoneme error rate over low resource languages and no degradation over high resource ones compared to monolingual baselines.

We present an end-of-utterance detector for real-time automatic speech recognition in far-field scenarios. The proposed system consists of three components: a long short-term memory (LSTM) neural network trained on acoustic features, an LSTM trained on l-best recognition hypotheses of the automatic speech recognition (ASR) decoder, and a feedforward deep neural network (DNN) combining embeddings derived from both LSTMs with pause duration features from the ASR decoder. At inference time, lower and upper latency (pause duration) bounds act as safeguards. Within the latency bounds, the utterance end-point is triggered as soon as the DNN posterior reaches a tuned threshold. Our experimental evaluation is carried out on real recordings of natural human interactions with voice-controlled far-field devices. We show that the acoustic embeddings are the single most powerful feature and particularly suitable for cross-lingual applications. We furthermore show the benefit of ASR decoder features, especially as a low cost alternative to ASR hypothesis em-beddings.

In this work, we propose a classifier for distinguishing device-directed queries from background speech in the context of interactions with voice assistants.Applications include rejection of false wake-ups or unintended interactions as well as enabling wake-word free followup queries.Consider the example interaction: "Computer, play music", "Computer, reduce the volume".In this interaction, the user needs to repeat the wake-word (Computer) for the second query.To allow for more natural interactions, the device could immediately re-enter listening state after the first query (without wake-word repetition) and accept or reject a potential follow-up as device-directed or background speech.The proposed model consists of two long short-term memory (LSTM) neural networks trained on acoustic features and automatic speech recognition (ASR) 1-best hypotheses, respectively.A feed-forward deep neural network (DNN) is then trained to combine the acoustic and 1-best embeddings, derived from the LSTMs, with features from the ASR decoder.Experimental results show that ASR decoder, acoustic embeddings, and 1-best embeddings yield an equal-error-rate (EER) of 9.3 %, 10.9 % and 20.1 %, respectively.Combination of the features resulted in a 44 % relative improvement and a final EER of 5.2 %.

Statistical language models (LM) play a key role in Automatic Speech Recognition (ASR) systems used by conversational agents. These ASR systems should provide a high accuracy under a variety of speaking styles, domains, vocabulary and argots. In this paper, we present a DNN-based method to adapt the LM to each user-agent interaction based on generalized contextual information, by predicting an optimal, context-dependent set of LM interpolation weights. We show that this framework for contextual adaptation provides accuracy improvements under different possible mixture LM partitions that are relevant for both (1) Goal-oriented conversational agents where it's natural to partition the data by the requested application and for (2) Non-goal oriented conversational agents where the data can be partitioned using topic labels that come from predictions of a topic classifier. We obtain a relative WER improvement of 3% with a 1-pass decoding strategy and 6% in a 2-pass decoding framework, over an unadapted model. We also show up to a 15% relative improvement in recognizing named entities which is of significant value for conversational ASR systems.

Language models (LM) for interactive speech recognition systems are trained on large amounts of data and the model parameters are optimized on past user data. New application intents and interaction types are released for these systems over time, imposing challenges to adapt the LMs since the existing training data is no longer sufficient to model the future user interactions. It is unclear how to adapt LMs to new application intents without degrading the performance on existing applications. In this paper, we propose a solution to (a) estimate n-gram counts directly from the hand-written grammar for training LMs and (b) use constrained optimization to optimize the system parameters for future use cases, while not degrading the performance on past usage. We evaluated our approach on new applications intents for a personal assistant system and find that the adaptation improves the word error rate by up to 15% on new applications even when there is no adaptation data available for an application.

This article presents a whisper speech detector in the far-field domain. The proposed system consists of a long-short term memory (LSTM) neural network trained on log-filterbank energy (LFBE) acoustic features. This model is trained and evaluated on recordings of human interactions with voice-controlled, far-field devices in whisper and normal phonation modes. We compare multiple inference approaches for utterance-level classification by examining trajectories of the LSTM posteriors. In addition, we engineer a set of features based on the signal characteristics inherent to whisper speech, and evaluate their effectiveness in further separating whisper from normal speech. A benchmarking of these features using multilayer perceptrons (MLP) and LSTMs suggests that the proposed features, in combination with LFBE features, can help us further improve our classifiers. We prove that, with enough data, the LSTM model is indeed as capable of learning whisper characteristics from LFBE features alone compared to a simpler MLP model that uses both LFBE and features engineered for separating whisper and normal speech. In addition, we prove that the LSTM classifiers accuracy can be further improved with the incorporation of the proposed engineered features.

The structured language model (SLM) of [1] was one of the first to successfully integrate syntactic structure into language models. We extend the SLM framework in two new directions. First, we propose a new syntactic hierarchical interpolation that improves over previous approaches. Second, we develop a general information-theoretic algorithm for pruning the underlying Jelinek-Mercer interpolated LM used in [1], which substantially reduces the size of the LM, enabling us to train on large data. When combined with hill-climbing [2] the SLM is an accurate model, space-efficient and fast for rescoring large speech lattices. Experimental results on broadcast news demonstrate that the SLM outperforms a large 4-gram LM. 1.

Long-span language models, such as those involving syntactic dependencies, produce more coherent text than their n-gram counterparts. However, evaluating the large number of sentence-hypotheses in a packed representation such as an ASR lattice is intractable under such long-span models both during decoding and discriminative training. The accepted compromise is to rescore only the N-best hypotheses in the lattice using the long-span LM. We present discriminative hill climbing, an efficient and effective discriminative training procedure for long-span LMs based on a hill climbing rescoring algorithm [1]. We empirically demonstrate significant computational savings as well as error-rate reduction over N-best training methods in a state of the art ASR system for Broadcast News transcription.

Accurate estimates of language model parameters are critical for building quality text generation systems, such as automatic speech recognition. However, text training data for a domain of interest is often unavailable. Instead, we use semi-supervised model adaptation; parameters are estimated using both unlabeled in-domain data (raw speech audio) and labeled out of domain data (text.) In this work, we present a new semi-supervised language model adaptation procedure for Maximum Entropy models with n-gram features. We augment the conventional maximum likelihood training criterion on out-of-domain text data with an additional term to minimize conditional entropy on in-domain audio. Additionally, we demonstrate how to compute conditional entropy efficiently on speech lattices using first- and second-order expectation semirings. We demonstrate improvements in terms of word error rate over other adaptation techniques when adapting a maximum entropy language model from broadcast news to MIT lectures.

We describe a new approach for rescoring speech lattices - with long-span language models or wide-context acoustic models - that does not entail computationally intensive lattice expansion or limited rescoring of only an N-best list. We view the set of word-sequences in a lattice as a discrete space equipped with the edit-distance metric, and develop a hill climbing technique to start with, say, the 1-best hypothesis under the lattice-generating model(s) and iteratively search a local neighborhood for the highest-scoring hypothesis under the rescoring model(s); such neighborhoods are efficiently constructed via finite state techniques. We demonstrate empirically that to achieve the same reduction in error rate using a better estimated, higher order language model, our technique evaluates fewer utterance-length hypotheses than conventional N-best rescoring by two orders of magnitude. For the same number of hypotheses evaluated, our technique results in a significantly lower error rate.

Although research has previously been done on multilingual speech recognition, it has been found to be very difficult to improve over separately trained systems. The usual approach has been to use some kind of “universal phone set” that covers multiple languages. We report experiments on a different approach to multilingual speech recognition, in which the phone sets are entirely distinct but the model has parameters not tied to specific states that are shared across languages. We use a model called a “Subspace Gaussian Mixture Model” where states' distributions are Gaussian Mixture Models with a common structure, constrained to lie in a subspace of the total parameter space. The parameters that define this subspace can be shared across languages. We obtain substantial WER improvements with this approach, especially with very small amounts of in-language training data.

In this paper we present a novel approach for estimating feature-space maximum likelihood linear regression (fMLLR) transforms for full-covariance Gaussian models by directly maximizing the likelihood function by repeated line search in the direction of the gradient. We do this in a pre-transformed parameter space such that an approximation to the expected Hessian is proportional to the unit matrix. The proposed algorithm is as efficient or more efficient than standard approaches, and is more flexible because it can naturally be combined with sets of basis transforms and with full covariance and subspace precision and mean (SPAM) models.

Preparation of a lexicon for speech recognition systems can be a significant effort in languages where the written form is not exactly phonetic. On the other hand, in languages where the written form is quite phonetic, some common words are often mispronounced. In this paper, we use a combination of lexicon learning techniques to explore whether a lexicon can be learned when only a small lexicon is available for boot-strapping. We discover that for a phonetic language such as Spanish, it is possible to do that better than what is possible from generic rules or hand-crafted pronunciations. For a more complex language such as English, we find that it is still possible but with some loss of accuracy.

We describe an acoustic modeling approach in which all phonetic states share a common Gaussian Mixture Model structure, and the means and mixture weights vary in a subspace of the total parameter space. We call this a Subspace Gaussian Mixture Model (SGMM). Globally shared parameters define the subspace. This style of acoustic model allows for a much more compact representation and gives better results than a conventional modeling approach, particularly with smaller amounts of training data.

In this paper, we present a novel version of discriminative training for N-gram language models. Language models impose language specific constraints on the acoustic hypothesis and are crucial in discriminating between competing acoustic hypotheses. As reported in the literature, discriminative training of acoustic models has yielded significant improvements in the performance of a speech recognition system, however, discriminative training for N-gram language models (LMs) has not yielded the same impact. In this paper, we present three techniques to improve the discriminative training of LMs, namely updating the back-off probability of unseen events, normalization of the N-gram updates to ensure a probability distribution and a relative-entropy based global constraint on the N-gram probability updates. We also present a framework for discriminative adaptation of LMs to a new domain and compare it to existing linear interpolation methods. Results are reported on the Broadcast News and the MIT lecture corpora. A modest improvement of 0.2% absolute (on Broadcast News) and 0.3% absolute (on MIT lectures) was observed with discriminatively trained LMs over state-of-the-art systems.

In this paper, we propose a new method for detecting regions with out-of-vocabulary (OOV) words in the output of a large vocabulary continuous speech recognition (LVCSR) system. The proposed method uses a hybrid system combining words and data-driven variable length sub word units. With the use of a single feature, the posterior probability of sub word units, this method outperforms existing methods published in the literature. We also presents a recipe to discriminatively train a hybrid language model to improve OOV detection rate. Results are presented on the RT04 broadcast news task.

This paper presents the advantages of augmenting a word-based system with sub-word units as a step towards building open vocabulary speech recognition systems. We show that a hybrid system which combines words and data-driven, variable length sub word units has a better phone accuracy than word only systems. In addition the hybrid system is better in detecting Out-Of-Vocabulary (OOV) terms and representing them phonetically. Results are presented on the RT-04 broadcast news and MIT Lecture data sets. An FSM-based approach to recover OOV words from the hybrid lattices is also presented. At an OOV rate of 2.5 % on RT-04 we observed a 8 % relative improvement in phone error rate (PER), 7.3 % relative improvement in oracle PER and 7 % relative improvement in WER after recovering the OOV terms. A significant reduction of 33 % relative in PER is seen in the OOV regions. Index Terms: hybrid LVCSR system, out-of-vocabulary, subword unit selection, phone recognition, OOV detection

This paper addresses the adjustment of the language model (LM) scaling factor of an automatic speech recognition (ASR) system for a new domain using only un-transcribed speech. The main idea is to replace the (unavailable) reference transcript with an automatic transcript generated by an independent ASR system, and adjust parameters using this sloppy reference. It is shown that despite its fairly high error rate (ca. 35%), choosing the scaling factor to minimize disagreement with the erroneous transcripts is still an effective recipe for model selection. This effectiveness is demonstrated by adjusting an ASR system trained on Broadcast News to transcribe the MIT Lectures corpus. An ASR system for telephone speech produces the sloppy reference, and optimizing towards it yields a nearly optimal LM scaling factor for the MIT Lectures corpus. Index Terms: LVCSR, language modeling, domain adaptation, semi-supervised learning, unsupervised learning

This paper addresses the detection of OOV segments in the output of a large vocabulary continuous speech recognition (LVCSR) system. First, standard confidence measures from frame-based wordand phone- posteriors are investigated. Substantial improvement is obtained when posteriors from two systems — strongly constrained (LVCSR) and weakly constrained (phone posterior estimator) are combined. We show that this approach is also suitable for detection of general recognition errors. All results are presented on WSJ task with reduced recognition vocabulary.