CHANNEL COMPENSATION OF TELEPHONE SPEECH RECOGNITION BASED ON MFCCs FILTERING

One of main tasks in robust telephone speech recognition is to overcome the channel noise. RASTA-PLP technique is an effective channel compensation method and has been widely discussed. In this paper,a new MFCCs channel compensation method is proposed, which has a functionality in both good performance and simple computation. At first,it uses Mel spectral analysis instead of PLP spectral analysis to reduce the computation. Based on the linear relationship between Mel-frequency log spectrum and MFCCs,it then extend RASTA processing from Mel-frequency log spectrum to MFCCs,and a RASTA-like band-pass filter is proposed for robut telephone speech recognition. Next, this paper selects the pole parameter of the filter by experiments and discusses the selection of the initial value.Comparing with RASTA-PLP processing, the proposed RMFCC method reduces the computational complexity without losing performance, and it is also better than CMS and two level CMS on the performance. After discussion, it is found that many channel compensation methods are based on the filtering processing ,and it is verified that it is an effective way to suppress very low modulation frequencies for bust telephone speech recognition.