Model-based Speech Enhancement With Multi-stream Technique

The noise is everywhere in people’s living environment, so the clean speech signal will be disturbed by noise easily. If we don’t process the noise, the speech quality will be worse and lead to auditory fatigue. So before delivering the speech signal to listener, it’s necessary to do speech enhancement so that eliminate the noise in speech signal. In recent years, model-based speech enhancement algorithm become an important research direction in speech enhancement domain, however, the speech signal analyzed in these algorithm is collected by microphone, so the characteristic of the input signal is single which result in limited capacity of speech enhancement. For the shortage, we proposed a model-based and multi-stream speech enhancement algorithm, by raising another data stream that collected by throat microphone, we obtain new characteristic of the signal that promote the performance of algorithm.In this dissertation, we combine the multi-stream technology and model-based algorithm, then propose a GMM-based and multi-stream speech enhancement algorithm. The work include the following aspects:1.Introduce several classical single channel algorithms and evaluate the results in chapter 4;2.Introduce GMM, multi-stream technology and how to apply them to the speech enhancement algorithm.3.Propose an algorithm that combine multi-stream and model-based algorithm, and construct the enhancement system with C++. First of all, we train clean speech GMM and noise GMM with HTK tools. Clean speech include air-conduction speech signal that is collected by microphone and non-air-conduction speech signal that is collected by throat microphone. Then calculate the noisy multi-stream GMM. At last, we do enhancement with prior information of models. With the evaluation of enhanced speech, it shows that the performance of combing multi-stream and model-based algorithm method is better than single-stream algorithm.