Abstract:
Automatic speech recognition (ASR) performance degrades in noisy environments such as in cars and factories. To improve ASR performance in various noise environments, microphone array noise reduction based on Wiener filtering was investigated. Although this is an effective method, several microphone array noise reduction parameters (noise reduction parameter set) are required, e.g., the flooring gain and stationary noise subtraction coefficient. Because the output sound quality varies with the noise reduction parameter set, ASR performance may degrade without optimal noise reduction parameter set selection. We propose a method of selecting the optimal noise reduction parameter set that will result in the highest character accuracy (CAcc) of ASR among multiple pre-adjusted noise reduction parameter sets. Through experiments using data recorded in running cars, CAcc improved to 68% with the proposed method.