Reasearch On The Methods Of Sound Dereverberation

When a microphone is located from a speaker too far away in some close room very narrow such as a cabin or a bathroom, sound signals it receives will be deteriorated by reverberation; in underwater acoustics communication, a similar phenomenon happens due to multi-path propagation; in applications of active sonar and sound bait, the interference from reverberation in fact is also one of the most crucial influence in addition to the environmental noise and self-noise. In order to improve the quality of the signals received, dereverberation technology is indispensably needed. This technology may also find applications in various aspects like seismic data analysis, ultrasonic measure, biomedicine, vibration acoustics etc. The research on dereverberation technology not only provides batter technique for these applications, but may also contribute to signal processing theoretically.So far, many kinds of dereverberation methods have been developed, such as complex cepstrum filtering, inverse filtering, beam forming and so on. In this thesis, dereverberation has been implemented through the strategies of 'avoiding reverberation' and 'reducing reverberation', coping with the four phases from sound to signal, i.e. the signal source, channel, signal picking and signal processing; and has surveyed many kinds of dereverberation methods. To overcome the drawbacks of complex cepstrum filtering and inverse filtering, the methods of complex cepstrum filtering with Gauss window and Wiener filtering with VAD have been explored; the evaluation criterions of two subjective methods, one objective method in time domain and three objective methods in frequency domain have been applied for speech quality improvements.Main work of the thesis includes:1. A survey of dereverberation related theories and methods. It has been discussed that the principle of room and ocean reverberation with the main parameters, their properties and the mathematical models; the definition and properties of complex cepstrum and the dereverberation principle; several kinds of deconvolution and VAD methods. Particularly, Wiener filtering arithmetic and one VAD arithmetic using double threshold zero-cross rate for dereverberation have been detailed in principle.2. A numerical simulation for dereverberation with complex cepstrum filtering technique of single channel. Respective to several evaluation criterions, the cut-off points of the lowpass filter in complex cepstral domain, the transition bandwidth and the curve characteristics in transition section have been obtained. It shows that, the cut-off point of the lowpass filter is independent of the reverberation time in ordinary reverberation time, and the dereverberation can be improved if a Gauss window is used before applying the complex cepstral filtering in time domain.3. Simulation investigations of dereverberation for speech signal from strong reverberation environments and the performance evaluation. A novel technique to obtain the room impulse response function, making use of the greeting words in regular speech commutation, has been proposed and verified. The deconvolution has been carried out with the obtained room response through Wiener filtering for speech samples from speech library. The dereverberation result judged by subjective feeling on base of word by word processing is satisfactory. With appropriate orders of Wiener filter, implementation of deconvolution for each word needs time in average about 0.3s to 0.5s on a personal computer.