Y. Ge, J. Niu (PRC), L. Ge, and K. Shirai (Japan)
Speech Recognition, Robustness, Non stationary measure, Feature with confident weights
The robustness of acoustic speech recognition system under noisy environment has become one of research focuses. One of key attempts to raise the robustness is to discover more available acoustic speech features to be able to describe the following signal natural character: a continuously random process repeatedly taking turn between stable pieces and non-stable pieces. Hence this paper suggests add several features to describe a trend degree and non-stable measure based on statistical theory. Meanwhile the paper introduces non-lineal random model RCA (Random Coefficient Auto-regression) to enhance ability dealing with non-linear signals. The other attempt is to improve Missing Feature Approach (MFA), which has been proved an available and considerable solution for robust speech recognition. However, MFA roughly classifies subband spectral components in binary way, reliable or unreliable, and is back strengths in deal with useful cepstral feature of speech. Based on analyses of three reliability standards, Feature with Confident Weights (FCW) and TC(Total Cepstral)-FCW are suggested by this paper. Our experimental results show that proposed features and approaches improve the recognition accuracy significantly in adverse environment, including stationary and non stationary noise environments.
Important Links:
Go Back
