Automatic Speech Recognition: A survey of deep learning techniques and approaches

Abstract

Significant research has been conducted during the last decade on the application of machine learning for speech processing, particularly speech recognition. However, in recent years, deep learning models have shown promising results for different speech related applications. With the emergence of end-to-end models, deep learning has revolutionized the field of Automatic Speech Recognition (ASR). A recent surge in transfer learning-based models and attention-based approaches on large datasets has further given an impetus to ASR. This paper provides a thorough review of the numerous studies conducted since 2010, as well as an extensive comparison of the state-of-the-art methods that are now being used in this research area, with a special focus on the numerous deep learning models, along with an analysis of contemporary approaches for both monolingual and multilingual models. Deep learning approaches are data dependent and their accuracy varies on different datasets. In this paper, we have also analysed the various models on publicly accessible speech datasets to understand model performance across diverse datasets for practical deployment. This study also highlights the research findings and challenges with way forward that may be used as a beginning point for academicians interested in open-source Automatic Speech Recognition (ASR) research, particularly focusing on mitigating data dependency and generalizability across low resource languages, speaker variability, and noise conditions. • Machine learning and deep learning models have transformed speech processing and recognition, with deep learning models showing promising results in recent years. • The emergence of end-to-end models, Transfer learning-based models and attention-based approaches, coupled with large datasets, have further enhanced Automatic Speech Recognition (ASR) techniques and performance. • The paper provides a comprehensive review and comparison of state-of-the-art methods in ASR, with a focus on deep learning models and both monolingual and multilingual approaches. • The study analyzes the performance of different models on publicly accessible speech datasets, highlighting the data dependency and variability in accuracy among deep learning approaches. • The study also identifies research findings, challenges, and future directions for open-source ASR research.

Topics & Concepts