Robust Speech Recognition Based on AMS Spectrum Subtraction and Confidence Interval Test

Amplitude Modulation spectrogram (AMS) can reflect the orthogonality between the "periodotopical" organization and the tonotopical organization in neurons. So we can use it as a reliable discrimination between speech and noise. This paper presents a novel approach of robust speech recognition by using AMS and Spectrum Subtraction with Confidence Interval Test. By dynamic using Confidence Interval Test, the error of estimated SNR was decreased. Experiment results show the proposed method yield good performance for attenuating the influence of noises and can improve the robustness of ASR.