Convolutional Neural Networks for Speech Recognition

Advancing Speech Recognition with CNNs

This research paper explores the application of Convolutional Neural Networks (CNNs) to enhance Automatic Speech Recognition (ASR) systems. The study highlights how CNNs, particularly with a novel limited weight sharing (LWS) scheme, can significantly improve performance over traditional Deep Neural Networks (DNNs) and Gaussian Mixture Model-Hidden Markov Models (GMM-HMMs).

The paper details the architecture of CNNs and their adaptation for speech processing, emphasizing features like local connectivity, weight sharing, and pooling. These characteristics contribute to invariance against variations in speech features, such as those caused by different speakers or environmental conditions.

Experimental results demonstrate that CNNs achieve a notable reduction in error rates, ranging from 6% to 10%, when compared to DNNs. These improvements were observed across standard benchmarks like the TIMIT phone recognition task and large-vocabulary voice search applications.

The research was conducted by authors affiliated with York University, the University of Toronto, and Microsoft Research, and was published in the IEEE/ACM Transactions on Audio, Speech, and Language Processing.