A study on robustness of large vocabulary mandarin Chinese continuous speech recognition system based on wavelet analysis

Abstract

In this paper wavelet decomposition is used to decompose speech signal into five levels. Low-frequency part of the speech signal was reconstructed. Because different frequencies of the speech signal have different influence on the performance of the system, the acoustic model of each level was trained and tested. The experimental results show that the acoustic model of level 1 is the best for clean speech and the acoustic model of level 2 is the best for noisy speech.It is proved that the frequency band of A1 makes a lot of contribution on the performance of clean speech and the frequency band of A2 makes a lot of contribution on the performance of noisy speech. © Springer-Verlag Berlin Heidelberg 2005.