The Speech Enhancement System

Speech enhancement is extensively used in communication systems and other areas. This study focuses on the problem of intelligibility enhancement and some new results are reported. For preparing to develop ASIC with own IP, the new algorithm to improve intelligibility has been realized by DSP. Six main results of the study are listed below: (1) A modified SNR algorithm SNRH for speech enhancement is suggested. Experiments show the new method is more related with intelligibility and outperforms the traditional SNR as it can offer more information regarding the applied conditions of enhancement approaches and their relative efficiency. (2) The effect of peak clipping on weighted cepstral coefficients is studied. It proves that peak clipping results in increase of relative small resonant peaks while keep the pattern of resonant peaks unchanged. Therefore, intelligibility decreases little after clipping. Greatly reduction of intelligibility in heavy noisy is mainly caused by the change of cross-zero time. (3) The effect of white noise on weighted cepstral coefficients indicates that unvoiced speech is more robust than voiced speech at same SNR. Randomness of white noise can not be neglected in low SNR. Unvoiced speech can be replaced by white noise without great change in ceptral distance. (4) A new algorithm is suggested to improve intelligibility by replacing non-speech segments with amplitude tunable white noise. Experiments results prove that this algorithm can dramatically improve intelligibility even the input SNR is as low as - 10dB. However other speech enhancement approaches improve SNR but decrease intelligibility in noisy environment. (5) The new algorithm is realized with hardware by TMS32OC541 fixed-point DSP of TI for further applications and IC design. The consumed memory are 2796 words of program, 852 words table ROM and 2162 words of RAM. The speeds are 5.8MIPS at initialization, 14.7MIPS/frame in beginning noise estimate, 39.8MIPS/frame to treat speech and 21.2MIPS/frame to deal with non-speech. In all case, the DSP can complete enhancement work in real time. (6) A cascade noisy resistant and interrupt resistant vocoder is developed by combination of the new algorithm and the phonetic classified vocoder. It uses enhanced speech signals to code and the bits of classification are protected by repeated parity check bits. The new vocoder can be used in many applications such as safety net-phones.