Large self-supervised models have excelled in various speech processing tasks, but their deployment on resource-limited devices is often impractical due to their substantial memory footprint. Previous studies have demonstrated the effectiveness of self-supervised pre-training for keyword spotting, even with constrained model capacity. In our pursuit of maintaining high performance while minimizing the model’s resource demands, we investigate the implementation of Quantization Aware Training for both self-supervised pre-training and fine-tuning, specifically tailored to fit within the constraints of on-device model budget. Our experiments emphasize the critical role of selecting and synchronizing QAT methods throughout both stages of model training and tuning. We evaluate our methodology on a 16.6k-hour in-house keyword spotting dataset, and show that there is no decline in performance, even when the bit size of model weights and activations is cut by a factor of four.

In this work, we present Slimmable Neural Networks applied to the problem of small-footprint keyword spotting. We show that slimmable neural networks allow us to create super-nets from Convolutional Neural Networks and Transformers, from which sub-networks of different sizes can be extracted. We demonstrate the usefulness of these models on in-house voice assistant data and Google Speech Commands, and focus our efforts on models for the on-device use case, limiting ourselves to less than 250k parameters. We show that slimmable models can match (and in some cases, outperform) models trained from scratch. Slimmable neural networks are therefore a class of models particularly useful when the same functionality is to be replicated at different memory and compute budgets, with different accuracy requirements.

Self-supervised speech representation learning (S3RL) is revolutionizing the way we leverage the ever-growing availability of data. While S3RL related studies typically use large models, we employ light-weight networks to comply with tight memory of compute-constrained devices. We demonstrate the effectiveness of S3RL on a keyword-spotting (KS) problem by using transformers with 330k parameters and propose a mechanism to enhance utterance-wise distinction, which proves crucial for improving performance on classification tasks. On the Google speech commands v2 dataset, the proposed method applied to the Auto-Regressive Predictive Coding S3RL led to a 1.2% accuracy improvement compared to training from scratch. On an in-house KS dataset with four different keywords, it provided 6% to 23.7% relative false accept improvement at fixed false reject rate. We argue this demonstrates the applicability of S3RL approaches to light-weight models for KS and confirms S3RL is a powerful alternative to traditional supervised learning for resource-constrained applications.

Fixed-point (FXP) inference has proven suitable for embedded devices with limited computational resources, and yet model training is continually performed in floating-point (FLP). FXP training has not been fully explored and the non-trivial conversion from FLP to FXP presents unavoidable performance drop. We propose a novel method to train and obtain FXP convolutional keyword-spotting (KWS) models. We combine our methodology with two quantization-aware-training (QAT) techniques – squashed weight distribution and absolute cosine regularization for model parameters, and propose techniques for extending QAT over transient variables, otherwise neglected by previous paradigms. Experimental results on the Google Speech Commands v2 dataset show that we can reduce model precision up to 4-bit with no loss in accuracy. Furthermore, on an in-house KWS dataset, we show that our 8bit FXP-QAT models have a 4-6% improvement in relative false discovery rate at fixed false reject rate compared to full precision FLP models. During inference we argue that FXP-QAT eliminates q-format normalization and enables the use of low-bit accumulators while maximizing SIMD throughput to reduce user-perceived latency. We demonstrate that we can reduce execution time by 68% without compromising KWS model’s predictive performance or requiring model architectural changes. Our work provides novel findings that aid future research in this area and enable accurate and efficient models.

We live in a world where 60% of the population can speak two or more languages fluently. Members of these communities constantly switch between languages when having a conversation. As automatic speech recognition (ASR) systems are being deployed to the real-world, there is a need for practical systems that can handle multiple languages both within an utterance or across utterances. In this paper, we present an end-to-end ASR system using a transformer-transducer model architecture for code-switched speech recognition. We propose three modifications over the vanilla model in order to handle various aspects of code-switching. First, we introduce two auxiliary loss functions to handle the low-resource scenario of code-switching. Second, we propose a novel mask-based training strategy with language ID information to improve the label encoder training towards intra-sentential code-switching. Finally, we propose a multi-label/multi-audio encoder structure to leverage the vast monolingual speech corpora towards code-switching. We demonstrate the efficacy of our proposed approaches on the SEAME dataset, a public Mandarin-English code-switching corpus, achieving a mixed error rate of 18.5% and 26.3% on testmanand testsgesets respectively.

We propose a novel approach to semi-supervised automatic speech recognition (ASR). We first exploit a large amount of unlabeled audio data via representation learning, where we reconstruct a temporal slice of filterbank features from past and future context frames. The resulting deep contextualized acoustic representations (DeCoAR) are then used to train a CTC-based end-to-end ASR system using a smaller amount of labeled audio data. In our experiments, we show that systems trained on DeCoAR consistently outperform ones trained on conventional filterbank features, giving 42% and 19% relative improvement over the baseline on WSJ eval92 and LibriSpeech test-clean, respectively. Our approach can drastically reduce the amount of labeled data required; unsupervised training on LibriSpeech then supervision with 100 hours of labeled data achieves performance on par with training on all 960 hours directly.

Voice-controlled house-hold devices, like Amazon Echo or Google Home, face the problem of performing speech recognition of device-directed speech in the presence of interfering background speech, i.e., background noise and interfering speech from another person or media device in proximity need to be ignored. We propose two end-to-end models to tackle this problem with information extracted from the anchored segment. The anchored segment refers to the wake-up word part of an audio stream, which contains valuable speaker information that can be used to suppress interfering speech and background noise. The first method is called Multi-source Attention where the attention mechanism takes both the speaker information and decoder state into consideration. The second method directly learns a frame-level mask on top of the encoder output. We also explore a multi-task learning setup where we use the ground truth of the mask to guide the learner. Given that audio data with interfering speech is rare in our training data, we also propose a way to synthesize "noisy" speech from "clean" speech to mitigate the mismatch between training and test data. Our proposed methods show up to 15% relative reduction in WER for Amazon Alexa live data with interfering background speech without significantly degrading on clean speech.

In this paper, we investigate how to apply graph-based semisupervised learning to acoustic modeling in speech recognition. Graph-based semisupervised learning is a widely used transductive semisupervised learning method in which labeled and unlabeled data are jointly represented as a weighted graph; the resulting graph structure is then used as a constraint during the classification of unlabeled data points. We investigate suitable graph-based learning algorithms for speech data and evaluate two different frameworks for integrating graph-based learning into state-of-the-art, deep neural network (DDN)-based speech recognition systems. The first framework utilizes graph-based learning in parallel with a DNN classifier within a lattice-rescoring framework, whereas the second framework relies on an embedding of graph neighborhood information into continuous space using an autoencoder. We demonstrate significant improvements in framelevel phonetic classification accuracy and consistent reductions in word error rate on large-vocabulary conversational speech recognition tasks.

Graph-based learning (GBL) is a form of semi-supervised learning that has been successfully exploited in acoustic modeling in the past. It utilizes manifold information in speech data that is represented as a joint similarity graph over training and test samples. Typically, GBL is used at the output level of an acoustic classifier; however, this setup is difficult to scale to large data sets, and the graph-based learner is not optimized jointly with other components of the speech recognition system. In this paper we explore a different approach where the similarity graph is first embedded into continuous space using a neural autoencoder. Features derived from this encoding are then used at the input level to a standard DNN-based speech recognizer. We demonstrate improved scalability and performance compared to the standard GBL approach as well as significant improvements in word error rate on a medium-vocabulary Switchboard task.

In this paper, we introduce a set of benchmark corpora of conversational English speech derived from the Switchboard-I and Fisher datasets. Traditional ASR research requires considerable computational resources and has slow experimental turnaround times. Our goal is to introduce these new datasets to researchers in the ASR and machine learning communities (especially in academia), in order to facilitate the development of novel acoustic modeling techniques on smaller but acoustically rich corpora. We select these corpora to maximize an acoustic quality criterion while limiting the vocabulary size (from 10 words up to 10,000 words) with different state-of-the-art submodular function optimization algorithms. We provide baseline word recognition results for both GMM and DNN-based systems and release the corpora definitions and Kaldi training recipes to the public.

We conduct a comparative study on selecting subsets of acoustic data for training phone recognizers. The data selection problem is approached as a constrained submodular optimization problem. Previous applications of this approach required transcriptions or acoustic models trained in a supervised way. In this paper we develop and evaluate a novel and entirely unsupervised approach, and apply it to TIMIT data. Results show that our method consistently outperforms a number of baseline methods while being computationally very efficient and requiring no labeling.

We address the problem of subselecting a large set of acoustic data to train automatic speech recognition (ASR) systems. To this end, we apply a novel data selection technique based on constrained submodular function maximization. Though NP-hard, the combinatorial optimization problem can be approximately solved by a simple and scalable greedy algorithm with constant-factor guarantees. We evaluate our approach by subselecting data from 1300 hours of conversational English telephone data to train two types large-vocabulary speech recognizers, one with Gaussian mixture model (GMM) based acoustic models, and another based on deep neural networks (DNNs). We show that training data can be reduced significantly, and that our technique outperforms both random selection and a previously proposed selection method utilizing comparable resources. Notably, using the submodular selection method, the DNN system using only about 5% of the training data is able to achieve performance on par with the GMM system using 100% of the training data - with the baseline subset selection methods, however, the DNN system is unable to accomplish this correspondence.

This paper describes the combination of two recent machine learning techniques for acoustic modeling in speech recognition: deep neural networks (DNNs) and graph-based semi-supervised learning (SSL). While DNNs have been shown to be powerful supervised classifiers and have achieved considerable success in speech recognition, graph-based SSL can exploit valuable complementary information derived from the manifold structure of the unlabeled test data. Previous work on graph-based SSL in acoustic modeling has been limited to frame-level classification tasks and has not been compared to, or integrated with, state-of-the-art DNN/HMM recognizers. This paper represents the first integration of graph-based SSL with DNN based speech recognition and analyzes its effect on word recognition performance. The approach is evaluated on two small vocabulary speech recognition tasks and shows a significant improvement in HMM state classification accuracy as well as a consistent reduction in word error rate over a state-of-the-art DNN/HMM baseline.

We apply methods for selecting subsets of dimensions from high-dimensional score spaces, and subsets of data for training, using submodular function optimization. Submodular functions provide theoretical performance guarantees while simultaneously retaining extremely fast and scalable optimization via an accelerated greedy algorithm. We evaluate this approach on two applications: data subset selection for phone recognizer training, and semi-supervised learning for phone segment classification. Interestingly, the first application uses submodularity twice: first for score space sub-selection and then for data subset selection. Our approach is computationally efficient but still consistently outperforms a number of baseline methods.

We present our system for the Interspeech 2013 Computational Paralinguistics Autism Sub-challenge. Our contribution focuses on improving classification accuracy of developmental disorders by applying a novel feature selection technique to the rich set of acoustic-prosodic features provided for this purpose. Our feature selection approach is based on submodular function optimization. We demonstrate significant improvements over systems using the full feature set and over a standard feature selection approach. Our final system outperforms the official Challenge baseline system significantly on the development set for both classification tasks, and on the test set for the Typicality task. Finally, we analyze the subselected features and identify the most important ones. Index Terms: classification, feature selection, neural networks, submodular functions

This paper presents several novel contributions to the emerging framework of graph-based semi-supervised learning for speech processing. First, we apply graph-based learning to variable-length segments rather than to the fixed-length vector representations that have been used previously. As part of this work we compare various graph-based learners, and we utilize an efficient feature selection technique for high-dimensional feature spaces that alleviates computational costs and improves the per-formance of graph-based learners. Finally, we present a method to improve regularization during the learning pro-cess. Experimental evaluation on the TIMIT frame and segment classification tasks demonstrates that the graph-based classifiers outperform standard baseline classifiers; furthermore, we find that the best learning algorithms are those that can incorporate prior knowledge. 1.

Several motor related Brain Computer Interfaces (BCIs) have been developed over the years that use activity decoded from the contralateral hemisphere to operate devices. Many recent studies have also talked about the importance of ipsilateral activity in planning of motor movements. For successful upper limb BCIs, it is important to decode finger movements from brain activity. This study uses ipsilateral cortical signals from humans (using ECoG) to decode finger movements. We demonstrate, for the first time, successful finger movement detection using machine learning algorithms. Our results show high decoding accuracies in all cases which are always above chance. We also show that significant accuracies can be achieved with the use of only a fraction of all the features recorded and that these core features also make sense physiologically. The results of this study have a great potential in the emerging world of motor neuroprosthetics and other BCIs.