Matched filtering approach to robust speech recognition

Abstract

Robustness of the performance of the automatic speech-recognition (ASR) systems has become important because of the widespread deployment of ASR in various information technology applications. This paper addresses robustness to environment noise in the speech signal. The speech pattern matching is recast as a sequence of sub-pattern matching problems in the time-frequency domain. Each sub-pattern matching is formulated as a 2D matched filter, which is known to be an optimum detector. This frequential detection is shown to provide robust recognition of the overall pattern. The new approach to ASR is evaluated on a limited vocabulary. speaker-dependent, isolated word-recognition task in an automobile acoustic environment and the results are promising.