Research On Adaptive Methods For Text-independent Speaker Recognition

Speaker recognition is one of the important research branches of speech signal processing field, which is an indispensable technique for establishing an information and harmonious society. However in the factual application, the laboratory"success"speaker recognition system can not satisfy the practical requirement in robustness, brainwave and adaptation ability and it becomes the bottleneck in the application of speaker recognition system.Speaker adaptation can minimize the distortion generated by various kinds of noises, channels and acoustic characteristics owing to the mismatch between the acoustic models and testing speech and increase the recognition rate, so it is always one of the kernel techniques and research hotspots in speaker recognition technique. The aim of this paper is on the basis of promoting the system's adaptation ability and practicability, involving the robustness analysis of correlated factors between speakers and environment and the research of relevant adaptation techniques .To sum up, they can be generalized as follows:(1) In the selection of feature parameters, with comparison of the different features between recognition rate and robustness aspects, this paper tries its best to find the feature parameters that have both higher recognition rate and more powerful ability of anti-noisy, which can be regarded as the suitable parameters of our speaker recognition system;(2) In the construction of the speaker model, through contrast among different models, this paper adopts the model with higher recognition rate and powerful robustness to describe the speaker recognition system;(3) Meanwhile this paper emphasizes the adaptation in the space of feature parameters and model parameters respectively. The former obtains pure characteristics of speech signal through handling it recorded by microphone, in other words, it gets the feature parameters similar to the feature under the training environment; the latter, through the adjust of the model parameters, make the model further approach the characteristic of testing environment. Therefore the distortion generated by mismatch of environment and acoustics characteristics can be minimized;(4) Besides, this paper puts forward an integrated algorithm through analyzing the advantages and disadvantages of MAP and MLLR, two common speaker adaptation methods, it combines both them by introducing a simplified MLLR module to MAP module, so when there is a few data, MLLR is better; while the data increases, MAP has embodied more merits.This paper illustrates the contribution to adaptation ability made by different feature parameters, speaker models, adaptation techniques and the comparative analysis of the results, at last proposes the prospect of future work.