Recent works have shown that prompting large language models with audio encodings can unlock speech recognition capabilities. However, existing techniques do not scale efficiently, especially while handling long form streaming audio inputs — not only do they extrapolate poorly beyond the audio length seen during training, but they are also computationally inefficient due to the quadratic cost of attention.In this work, we introduce SpeechLLM-XL, a linear scaling decoder-only model for streaming speech recognition. We process audios in configurable chunks using limited attention window for reduced computation, and the text tokens for each audio chunk are generated auto-regressively until an EOS is predicted. During training, the transcript is segmented into chunks, using a CTC forced alignment estimated from encoder output. SpeechLLM-XL with 1.28 seconds chunk size achieves 2.7%/6.7% WER on LibriSpeech test clean/other, and it shows no quality degradation on long form utterances 10x longer than the training utterances.

Large language models (LLMs) have become proficient at solving a wide variety of tasks, including those involving multi-modal inputs. In particular, instantiating an LLM (such as LLaMA) with a speech encoder and training it on paired data imparts speech recognition (ASR) abilities to the decoder-only model, hence called Speech-LLaMA. Nevertheless, due to the sequential nature of auto-regressive inference and the relatively large decoder, Speech-LLaMA models require relatively high inference time. In this work, we propose to speed up Speech-LLaMA inference by predicting multiple tokens in the same decoding step. We explore several model architectures that enable this, and investigate their performance using threshold-based and verification-based inference strategies. We also propose a prefix-based beam search decoding method that allows efficient minimum word error rate (MWER) training for such models. We evaluate our models on a variety of public benchmarks, where they reduce the number of decoder calls by ∼3.2x while maintaining or improving WER performance.

Speech LLMs use speech embeddings as the prompt to a Large Language Model (LLM) and generate human readable text for the speech signal in an autoregressive manner. Teacher-forcing is a common approach used for training Speech LLMs, which is dissimilar to the procedure used during inference, creating a gap between training and inference known as exposure bias. To mitigate exposure bias, we propose Speech-N-LlaMA. Contrary to the existing Speech LLMs that have single pass through the LLM during training, Speech-N-LlaMA incorporates multi-pass training. Through multiple passes, Speech-N-LlaMA mitigates exposure bias and uses the error correction capability of LLM to improve the performance of Speech LLMs. We propose an N-pass loss and utterance level temperature sampling in Speech-N-LlaMA to achieve this. We evaluate four different model sizes on three benchmarks, and show up to 18% relative improvement in Word Error Rate (WER) compared to baseline Speech LLM, while not incurring any additional compute during inference.

The growing popularity of multi-channel wearable devices, such as smart glasses, has led to a surge of applications such as targeted speech recognition and enhanced hearing. However, current approaches to solve these tasks use independently trained models, which may not benefit from large amounts of unlabeled data. In this paper, we propose M-BEST-RQ, the first multi-channel speech foundation model for smart glasses, which is designed to leverage large-scale self-supervised learning (SSL) in an array-geometry agnostic approach. While prior work on multi-channel speech SSL only evaluated on simulated settings, we curate a suite of real downstream tasks to evaluate our model, namely (i) conversational automatic speech recognition (ASR), (ii) spherical active source localization, and (iii) glasses wearer voice activity detection, which are sourced from the MMCSG and EasyCom datasets. We show that a general-purpose M-BEST-RQ encoder is able to match or surpass supervised models across all tasks. For the conversational ASR task in particular, using only 8 hours of labeled speech, our model outperforms a supervised ASR baseline that is trained on 2000 hours of labeled data, which demonstrates the effectiveness of our approach.

We propose DOC-RAG - Domain-distributed Co-occurrence Retrieval Augmentation for ASR language model personalization aiming to improve the automatic speech recognition of rare word patterns in unseen domains. Our approach involves contrastively training a document retrieval module to rank external knowledge domains based on their semantic similarity with respect to the input query. We further use n-gram co-occurrence distribution to recognize rare word patterns associated with specific domains. We aggregate the next word probability distribution based on the relative importance of different domains. Extensive experiments on three user-specific speech-to-text tasks for meetings, TED talks, and financial earnings calls show that DOC-RAG significantly outperforms strong baselines with an 8-15% improvement in terms of perplexity and a 4-7% reduction in terms of Word Error Rates in various settings.

ASR models are commonly trained with the cross-entropy criterion to increase the probability of a target token sequence. While optimizing the probability of all tokens in the target sequence is sensible, one may want to de-emphasize tokens that reflect transcription errors. In this work, we propose a novel token-weighted RNN-T criterion that augments the RNN-T objective with token-specific weights. The new objective is used for mitigating accuracy loss from transcriptions errors in the training data, which naturally appear in two settings: pseudo-labeling and human annotation errors. Experiments results show that using our method for semi-supervised learning with pseudo-labels leads to a consistent accuracy improvement, up to 38% relative. We also analyze the accuracy degradation resulting from different levels of WER in the reference transcription, and show that token-weighted RNN-T is suitable for overcoming this degradation, recovering 64% 99% of the accuracy loss.

Standard Recurrent Neural Network Transducers (RNN-T) decoding algorithms for speech recognition are iterating over the time axis, such that one time step is decoded before moving on to the next time step. Those algorithms result in a large number of calls to the joint network, which were shown in previous work to be an important factor that reduces decoding speed. We present a decoding beam search algorithm that batches the joint network calls across a segment of time steps, which results in 20 %-96 % decoding speedups consistently across all models and settings experimented with. In addition, aggregating emission probabilities over a segment may be seen as a better approximation to finding the most likely model output, causing our algorithm to improve oracle word error rate by up to 11 % relative as the segment size increases, and to slightly improve general word error rate.

This paper introduces a fast-slow encoder based transducer with streaming deliberation for end-to-end automatic speech recognition. We aim to improve the recognition accuracy of the fast-slow encoder based transducer while keeping its latency low by integrating a streaming deliberation model. Specifically, the deliberation model leverages partial hypotheses from the streaming fast encoder and implicitly learns to correct recognition errors. We modify the parallel beam search algorithm for fast-slow encoder based transducer to be efficient and compatible with the deliberation model. In addition, the deliberation model is designed to process streaming data. To further improve the deliberation performance, a simple text augmentation approach is explored. We also compare LSTM and Conformer models for encoding partial hypotheses. Experiments on Librispeech and in-house data show relative WER reductions (WERRs) from 3% to 5% with a slight increase in model size and negligible extra token emission latency compared with fast-slow encoder based transducer. The deliberation model also reduces rare WERs by 3-7% on large-scale in-house data. Compared with vanilla neural transducers, the proposed deliberation model together with fast-slow encoder based transducer obtains relative 10-11% WERRs on Librispeech and around relative 6% WERR on in-house data with smaller emission delays.

With 4.5 million hours of English speech from 10 different sources across 120 countries and models of up to 10 billion parameters, we explore the frontiers of scale for automatic speech recognition.We propose data selection techniques to efficiently scale training data to find the most valuable samples in massive datasets.To efficiently scale model sizes, we leverage various optimizations such as sparse transducer loss and model sharding.By training 1-10B parameter universal English ASR models, we push the limits of speech recognition performance across many domains.Furthermore, our models learn powerful speech representations with zero and few-shot capabilities on novel domains and styles of speech, exceeding previous results across multiple in-house and public benchmarks.For speakers with disorders due to brain damage, our best zero-shot and few-shot models achieve 22% and 60% relative improvement on the AphasiaBank test set, respectively, while realizing the best performance on public social media videos.Furthermore, the same universal model reaches equivalent performance with 500x less in-domain data on the SPGISpeech financial-domain dataset.

Packet loss may affect a wide range of applications that use voice over IP (VoIP), e.g. video conferencing. In this paper, we investigate a time-domain convolutional recurrent network (CRN) for online packet loss concealment. The CRN comprises a convolutional encoder-decoder structure and long short-term memory (LSTM) layers, which have been shown to be suitable for real-time speech enhancement applications. Moreover, we propose lookahead and masked training to further improve the performance of the CRN framework. Experimental results show that the proposed system outperforms a baseline system using only LSTM layers in terms of two objective metrics – perceptual evaluation of speech quality (PESQ) and short-term objective intelligibility (STOI); it also reduces the word error rate (WER) more than the baseline when used as a frontend for speech recognition. The advantage of the proposed system is also verified in a subjective evaluation by the mean opinion score (MOS).

How to leverage dynamic contextual information in end-toend speech recognition has remained an active research area.Previous solutions to this problem were either designed for specialized use cases that did not generalize well to open-domain scenarios, did not scale to large biasing lists, or underperformed on rare long-tail words.We address these limitations by proposing a novel solution that combines shallow fusion, trie-based deep biasing, and neural network language model contextualization.These techniques result in significant 19.5% relative Word Error Rate improvement over existing contextual biasing approaches and 5.4%-9.3%improvement compared to a strong hybrid baseline on both open-domain and constrained contextualization tasks, where the targets consist of mostly rare long-tail words.Our final system remains lightweight and modular, allowing for quick modification without model re-training.

End-to-end models in general, and Recurrent Neural Network Transducer (RNN-T) in particular, have gained significant traction in the automatic speech recognition community in the last few years due to their simplicity, compactness, and excellent performance on generic transcription tasks. However, these models are more challenging to personalize compared to traditional hybrid systems due to the lack of external language models and difficulties in recognizing rare long-tail words, specifically entity names. In this work, we present novel techniques to improve RNN-T's ability to model rare WordPieces, infuse extra information into the encoder, enable the use of alternative graphemic pronunciations, and perform deep fusion with personalized language models for more robust biasing. We show that these combined techniques result in 15.4%-34.5% relative Word Error Rate improvement compared to a strong RNN-T baseline which uses shallow fusion and text-to-speech augmentation. Our work helps push the boundary of RNN-T personalization and close the gap with hybrid systems on use cases where biasing and entity recognition are crucial.

There is a growing interest in the speech community in developing Recurrent Neural Network Transducer (RNN-T) models for automatic speech recognition (ASR) applications. RNN-T is trained with a loss function that does not enforce temporal alignment of the training transcripts and audio. As a result, RNN-T models built with uni-directional long short term memory (LSTM) encoders tend to wait for longer spans of input audio, before streaming already decoded ASR tokens. In this work, we propose a modification to the RNN-T loss function and develop Alignment Restricted RNN-T (Ar-RNN-T) models, which utilize audio-text alignment in-formation to guide the loss computation. We compare the proposed method with existing works, such as monotonic RNN-T, on LibriSpeech and in-house datasets. We show that the Ar-RNN-T loss provides a refined control to navigate the trade-offs between the token emission delays and the Word Error Rate (WER). The Ar-RNN-T models also improve downstream applications such as the ASR End-pointing by guaranteeing token emissions within any given range of latency. Moreover, the Ar-RNN-T loss allows for bigger batch sizes and 4 times higher throughput for our LSTM model architecture, enabling faster training and convergence on GPUs.

Abstract The unprecedented growth of noise pollution over the last decades has raised an always increasing need for developing efficient audio enhancement technologies. Yet, the variety of difficulties related to processing audio sources in-the-wild, such as handling unseen noises or suppressing specific interferences, makes audio enhancement a still open challenge. In this regard, we present (the Neuro-Holistic Audio-eNhancement System), a Python toolkit for in-the-wild audio enhancement that includes functionalities for audio denoising, source separation, and —for the first time in such a toolkit—selective noise suppression. The architecture is specially developed to automatically adapt to different environmental backgrounds and speakers. This is achieved by the use of two identical neural networks comprised of stacks of residual blocks, each conditioned on additional speech- and noise-based recordings through auxiliary sub-networks. Along to a Python API, a command line interface is provided to researchers and developers, both of them carefully documented. Experimental results indicate that achieves great performance w. r. t. existing methods, preserving also the audio quality at a high level; thus, ensuring a reliable usage in real-life application, e. g., for in-the-wild speech processing, which encourages the development of speech-based intelligent technology.

End-to-end (E2E) systems for automatic speech recognition (ASR), such as RNN Transducer (RNN-T) and Listen-Attend-Spell (LAS) blend the individual components of a traditional hybrid ASR system - acoustic model, language model, pronunciation model - into a single neural network. While this has some nice advantages, it limits the system to be trained using only paired audio and text. Because of this, E2E models tend to have difficulties with correctly recognizing rare words that are not frequently seen during training, such as entity names. In this paper, we propose modifications to the RNN-T model that allow the model to utilize additional metadata text with the objective of improving performance on these named entity words. We evaluate our approach on an in-house dataset sampled from de-identified public social media videos, which represent an open domain ASR task. By using an attention model and a biasing model to leverage the contextual metadata that accompanies a video, we observe a relative improvement of about 16% in Word Error Rate on Named Entities (WER-NE) for videos with related metadata.

We consider neural network training, in applications in which there are many possible classes, but at test-time, the task is a binary classification task of determining whether the given example belongs to a specific class. We define the Single Logit Classification (SLC) task: training the network so that at test-time, it would be possible to accurately identify whether the example belongs to a given class in a computationally efficient manner, based only on the output logit for this class. We propose a natural principle, the Principle of Logit Separation, as a guideline for choosing and designing loss functions that are suitable for SLC. We show that the Principle of Logit Separation is a crucial ingredient for success in the SLC task, and that SLC results in considerable speedups when the number of classes is large.

The use of deep learning (DL) architectures for speech enhancement has recently improved the robustness of voice applications under diverse noise conditions.These improvements are usually evaluated based on the perceptual quality of the enhanced audio or on the performance of automatic speech recognition (ASR) systems.We are interested instead in the usefulness of these algorithms in the field of speech emotion recognition (SER), and specifically in whether an enhancement architecture can effectively remove noise while preserving enough information for an SER algorithm to accurately identify emotion in speech.We first show how a scalable DL architecture can be trained to enhance audio signals in a large number of unseen environments, and go on to show how that can benefit common SER pipelines in terms of noise robustness.Our results show that incorporating a speech enhancement architecture is beneficial, especially for low signal-to-noise ratio (SNR) conditions.

We consider neural network training, in applications in which there are many possible classes, but at test-time, the task is a binary classification task of determining whether the given example belongs to a specific class, where the class of interest can be different each time the classifier is applied. For instance, this is the case for real-time image search. We define the Single Logit Classification (SLC) task: training the network so that at test-time, it would be possible to accurately identify whether the example belongs to a given class in a computationally efficient manner, based only on the output logit for this class. We propose a natural principle, the Principle of Logit Separation, as a guideline for choosing and designing losses suitable for the SLC. We show that the cross-entropy loss function is not aligned with the Principle of Logit Separation. In contrast, there are known loss functions, as well as novel batch loss functions that we propose, which are aligned with this principle. In total, we study seven loss functions. Our experiments show that indeed in almost all cases, losses that are aligned with the Principle of Logit Separation obtain at least 20% relative accuracy improvement in the SLC task compared to losses that are not aligned with it, and sometimes considerably more. Furthermore, we show that fast SLC does not cause any drop in binary classification accuracy, compared to standard classification in which all logits are computed, and yields a speedup which grows with the number of classes. For instance, we demonstrate a 10x speedup when the number of classes is 400,000.

When humans learn a new concept, they might ignore examples that they cannot make sense of at first, and only later focus on such examples, when they are more useful for learning. We propose incorporating this idea of tunable sensitivity for hard examples in neural network learning, using a new generalization of the cross-entropy gradient step, which can be used in place of the gradient in any gradient-based training method. The generalized gradient is parameterized by a value that controls the sensitivity of the training process to harder training examples. We tested our method on several benchmark datasets. We propose, and corroborate in our experiments, that the optimal level of sensitivity to hard example is positively correlated with the depth of the network. Moreover, the test prediction error obtained by our method is generally lower than that of the vanilla cross-entropy gradient learner. We therefore conclude that tunable sensitivity can be helpful for neural network learning.

The adage that there is no data like more data is not new in affective computing; however, with recent advances in deep learning technologies, such as end-to-end learning, the need for extracting big data is greater than ever. Multimedia resources available on social media represent a wealth of data more than large enough to satisfy this need. However, an often prohibitive amount of effort has been required to source and label such instances. As a solution, we introduce Cost-efficient Audio-visual Acquisition via Social-media Small-world Targeting (CAS2T) for efficient large-scale big data collection from online social media platforms. Our system is based on a unique combination of small-world modelling, unsupervised audio analysis, and semi-supervised active learning. Such an approach facilitates rapid training on entirely new tasks sourced in their entirety from social multimedia. We demonstrate the high capability of our methodology via collection of original datasets containing a range of naturalistic, in-the-wild examples of human behaviours.

Dimensional emotion recognition from physiological signals is a highly challenging task. Common methods rely on hand-crafted features that do not yet provide the performance necessary for real-life application. In this work, we exploit a series of convolutional and recurrent neural networks to predict affect from physiological signals, such as electrocardiogram and electrodermal activity, directly from the raw time representation. The motivation behind this so-called end-to-end approach is that, ultimately, the network learns an intermediate representation of the physiological signals that better suits the task at hand. Experimental evaluations show that, this very first study on end-to-end learning of emotion based on physiology, yields significantly better performance in comparison to existing work on the challenging RECOLA database, which includes fully spontaneous affective behaviors displayed during naturalistic interactions. Furthermore, we gain better understanding of the models' inner representations, by demonstrating that some cells' activations in the convolutional network are correlated to a large extent with hand-crafted features.

Abstract We prove that all known examples of weakly o-minimal nonvaluational structures have no definable Skolem functions. We show, however, that such structures eliminate imaginaries up to definable families of cuts. Along the way we give some new examples of weakly o-minimal nonvaluational structures.

Traditional convolutional layers extract features from patches of data by applying a non-linearity on an affine function of the input. We propose a model that enhances this feature extraction process for the case of sequential data, by feeding patches of the data into a recurrent neural network and using the outputs or hidden states of the recurrent units to compute the extracted features. By doing so, we exploit the fact that a window containing a few frames of the sequential data is a sequence itself and this additional structure might encapsulate valuable information. In addition, we allow for more steps of computation in the feature extraction process, which is potentially beneficial as an affine function followed by a non-linearity can result in too simple features. Using our convolutional recurrent layers, we obtain an improvement in performance in two audio classification tasks, compared to traditional convolutional layers.

We use a feedforward Convolutional Neural Network to classify speakers' native language for the INTERSPEECH 2016 Computational Paralinguistic Challenge Native Language Sub-Challenge, using no specialized features for computational paralinguistics tasks, but only MFCCs with their first and second order deltas.In addition, we augment the training data by replacing the original examples with shorter overlapping samples extracted from them, thus multiplying the number of training examples by almost 40.With the augmented training dataset and enhancements to neural network models such as Batch Normalization, Dropout, and Maxout activation function, we managed to improve upon the challenge baseline by a large margin, both for the development and the test set.