Noise robust chinese speech recognition system for isolate words

Abstract

Summary form only given. A noise robust Chinese speech recognition system is built by appending an implementation of fundamental frequency (FF) estimation to a non-tonal language recognition system. Since FF detection is crucially important for the tone modeling of Chinese, and the widely used FF detection method, AUTOC, is vulnerable to a serious noise environment, the paper proposes a new algorithm, named running spectrum filtering (RSF), which is added to AUTOC to improve the anti-noise ability. Some new adjustments are also made to the traditional detection method. From these considerations, the errors in the FF contour caused by noise distortion are apparently amended. An evaluation experiment is undertaken using 12 Chinese words as the database to compare the proposed recognition system with the conventional system for noise levels of 10-20 dB SNR (white noise, pink noise and car interior noise); the results show that the recognition accuracy of the proposed system is significantly improved.