Speech feature smoothing for robust ASR

Abstract

Speech enhancement module is the key component in noise robust Automatic Speech Recognizer. A number of research studies have been conducted to improve the recognition accuracy of the Automatic Speech Recognizer. In this work, an attempt has been made to develop one such speech enhancement technique that is based on the existing popular technique called the Stereo-based Piecewise Linear Compensation for Environments. This report presents the review of the literature related to this work. It also describes the proposed methodology, and presents the results obtained using the same on the benchmark speech database called the Aurora2. The results show that the proposed algorithm is capable of achieving better accuracy than that obtained for basic Stereo-based Piecewise Linear Compensation for Environments algorithm.