Research On Computational Auditory Scene Analysis For Concurrent Speech Signals

The purpose of Computational Auditory Scene Analysis(CASA) is to modify the psychological and physiological function of human's auditory system by using computer technique, to make computer to have the ability to process sound like human ear do(including segregation and explanation). This is a brand new topic for research. In this dissertation, scene analysis for concurrent speech signals are discussed, and we have set up a basically complete concurret speech signals ASA system The novel achievements are listed as follows:1. For the bad separation results, complexity of grouping process and huge computation in the existing ASA system for concurrent speech signals, refering to Multi-Band Excitation (MBE) technique, we present a double pitch multi band excitation scene analysis model (DP-MBE SAM), and a new ASA system for concurrent speech signals based on DP-MBE SAM. The new system differs from the existing ASA system in that mixed signals is not decomposed into elements with fixed filters banks, but according to the given pitches, their parameters are estimated simultaneously and naturally into two speech groups, thus the DP-MBE SAM model is more applicable to the frequency change of the speech signals and possess the roubustness and flexibility in concurrent speech signals separation, whose character is close to human auditory system. The new system presents a more efficient way to decompose close harmonies than the existing ASA system and reduces the complexity of grouping. Besides, the mechanism of DP-MBE SAM can be extended to the scene analysis of more than two concurrent speech signals, and the parameters of the speech is a middle level present to communicate the low level auditory system and the high level brain. Experimental results show that concurrent speeches with different pitches are separated efficiently.2 . For the existing problem of the ASA system for concurrent speech signals based on DP-MBE SAM in chapter 3, we present an improved ASA system based on DP-MBE-SAM. The improvement include two parts: firstly, for the parameter ambiguity caused by the singularity of parameter extracting matrix, and other ambiguities, by referring to multi-frame interpolation method. We present an improved DP-MBE SAM; secondly, un-voice analysis is introduced to the DP-MBE-SAM, so that the improved ASA system can be used not only to separate concurrent vowels but concurrent speech signals with unvoice as well. Experimental results show the efficiency of the improved ASA system.3. For the existing problem of Meddis' psychophysically faithful method for extracting pitches from concurrent speech signalsis, we present a new pitch extracting method, which uses close-loop adaptive frequencies picking block to...