Study On Methods For Speech Enhancement Based On Dual Microphone Array

Under the real environment, speech be interfered by reverberation and noise, those interferences will reduce intelligibility and clarity of speech. Thus speech enhancement processing is necessary for noise speech. The main purpose of speech enhancement can be attributed to two aspects——to enhance the quality of speech and improve the intelligibility of speech, it is difficult to achieve both goals simultaneously. The main challenge of speech enhancement is to suppress noise in the signal without introducing distortion. With respect to the single microphone, the time domain information of the speech and airspace information of array can be used by the microphone array for speech enhancement, thus, higher quality enhanced speech can be obtained. However, the development trend of electronic equipment is miniaturized, the size of microphone array is limited. Herein, this dissertation focuses on dual microphone array based speech enhancement methods, and the contributions are as follows:The algorithm principle of fixed beamformer, generalized sidelobe canceller and post-filter is described in detail, and denoising performance of these algorithms in the model of dual microphone array is analyzed. Under different noise and SNR, denoising performance of these algorithms are tested through simulation.A speech enhancement algorithm which based on the combination of coherence filter and generalized sidelobe canceller is analyzed. In the algorithm, coherence filter is placed on the non-adaptive branch of generalized sidelobe canceller. This is a classic combination mode,and the analysis found that the coherence filter model is limited to this algorithm. To solve this problem, a new combined model is proposed. In new combined model, the coherence filter is a post filter of generalized sidelobe canceller. It overcome the limitations of coherence filter in the classical model, a improved coherence filter which has stronger denoising performance can be used in the new model.A coherence filter algorithm which based on iterative noise power spectral density estimation is analyzed, the analysis revealed that there is a problem of noise over-estimation. To solve this problem, an iterative noise power spectral density estimation that based on time smoothing parameter is proposed. This method applied a time-recursive averaging technique, and combined with iterative noise power spectral density estimation. Compared with the algorithm based on iterative noise power spectral density estimation, a more accurate estimation of the noise power spectral density can be obtained by this algorithm.The improved iterative noise power spectral density estimation technique is applied to coherence filter, and the coherence filter is combined with generalized sidelobe canceller, thus, a speech enhancement method based on the combination of generalized sidelobe canceller and coherence filter for dual microphone array is obtained. A large number of simulation experiments under different noise environments show that, compared with the speech enhancement algorithm that based on the classic combination model, higher quality enhanced speech can be obtained by proposed algorithm.