Joint compensation of noise and channel in speech recognition

A method named spectral and cepstral joint adjusting was proposed in order to jointly compensate for the channel mismatch and existence of additive noise in speech recognition.The method compensates for the noise and channel mismatch in spectral and cepstral domain respectively based on the ML rule.It didn't simplify the model of channel and noise effect on speech,and therefore reduce the error brought by the model simplification.Furthermore, the time complicacy was relatively low.The experiment on noisy Chinese digit string recognition with SNR of 10 dB to 20 dB was carried out.The results showed that the proposed method obtained the relative error rate reduction of 50.44%,which indicated that the spectral and cepstral joint adjusting method could rapidly compensate for the channel and noise well.