Implementation Technique And Method Of Blind Speech Separation For Convolutive Mixtures

Blind source separation(BSS) aims to recover original source signals only using observed signals, and mixing process is unknown or a little known. Blind Speech Separation is a challenging issue in BSS field, it will be used widely in speech enhancement, speech recognition, speaker location and high-quality hearing aids and currently it’s still in theory. The speech mixtures mainly includes instantaneous mixtures and convolutive mixtures. In real environment speech signals is of multipath effect and time delay, so BSS for convolutive mixture is closer to reality and is becoming a key issue.There are two major approaches to solving the blind speech separation for convolutive mixtures, the time domain approach and the frequency domain approach. The frequency domain is of low computational cost and good separation effect, where classic instantaneous BSS methods can be applied directly. But the permutation problem is the most critical issue. The convolutive BSS is more complex than the instantaneous BSS because of the mixing model, so currently most algorithms run in PC and embedded system solutions is rare. Technology products are becoming smarter and smaller, and the transplant to embedded platform is meaningful and significant.In response to the above issues, we research the method and hardware implementation of the blind speech separation for convolutive mixtures in a real environment. The paper proposes a frequency domain method where using a clustering to solve the permutation problem. The paper proposes and implements a blind speech separation system solution based on DSP TMS320C6713. Besides we also research how to design blind speech separation software application based on Windows OS using Windows multimedia API. The contribution of the paper contains:(1) Summarize the history of blind speech separation and current situation, and introduce BSS for convolutive mixtures basic principles. Meanwhile analyse and compare the advantages and disadvantages of time domain method and frequency domain method.(2) Propose a frequency domain approach to the convolutive blind speech separation, and the algorithm includes three steps. First the time domain signal is transform to the frequency domain using STFT, and use complex-valued JADE algorithm in every frequency bin. Then use a clustering method to solve the permutation problem and use the minimal distortion principle to solve amplitude problem. At last get the separation filter and filter the observed signals to get the separation signals. PC simulations show the algorithm has good separation performance.(3) Propose a embedded system solution for blind speech separation in real environment, and make a BSS embedded application primary exploration. The system chooses TI DSP TMS320C6713as the hardware core platform, and designs the front two channels microphone capture circuit based on the TMS320C6713DSK. We make full use of the hardware resources, and the blind speech separation on the DSP platform is implemented in CCS3.1. Meanwhile enough software optimizations ensure the good efficiency. The experimental results show good performance on two speakers’ mixture separation in a real room environment on both offline and online scenarios.(4) Propose a design of blind speech separation software application based on Windows OS. and peripheral hardware mainly contains a USB sound card and two condenser microphones. We use C/C++to develop the system in Visual C++6.0, and design a reasonable system framework. The software core contains three threads:the speech acquisition thread, the BSS algorithm thread and the speech play thread. We designs a simple and friendly UI based on the MFC and manipulate the audio data How using Windows multimedia API. The test results show good performance and stability on two speakers’mixture separation in a real environment.