The Research Of Speech Separation Based On Computational Auditory Scene Analysis

Computational Auditory Scene Analysis (CASA) is an emerging subject in the field of speech signal processing. In natural environment, multiple sound source speech signals form a typical auditory scene, the using of the Auditory Scene Analysis (ASA) can effectively separate target voice from complex acoustic environment and achieves a good performance. However, with the requirement of human-computer interaction, Automatic Speech Recognition (ASR) and hearing aid research, the speech separation algorithm in the noisy environment is still confronting with relatively big challenges.This paper mainly concentrates on not only researching and generalizing Computational Auditory Scene Analysis (CASA), but also analyzing the characteristics of current separation algorithm, sound and noise. In addition, it will pour more attention into the research of speech separations. Detailed work and innovation are presented as follows:Human auditory system deals with high and low frequency in different mechanisms and the mixed speech signal has low energy in high frequency, which is easier to be affected by noise. According to the property, this thesis proposes a speech separation algorithm based on signal energy and it can calculate energy before the auditory segmentation. Therefore, according to the ratio of energy removes signals in high frequency which is more likely to come from Time-Frequency units of the noise. In doing so, targeted speech signals will be less affected by noise, which generates more effective results.It is barely impossible for the signals of different sound source to have the same starting time and end time. Thus, to separate mixed speech signals can be completed by the sound cues. This paper adopts an accurate envelope extraction algorithm that can extract the signal onset and offset, then smooth signals, tests and combines selected onsets and offsets and obtain auditory segmentation, which finally separates targeted speech signals on the basis of binary mode kno wledge.The thesis chooses Cookie data sets to simulate experiment under Visual C++6.0and has evaluated experimental results, which verifies the feasibility and effectiveness of the proposed algorithm.