Research On Speech Recognition Algorithms And Its Implementation On DSP

With the advancement of the portable computers and the computing environments be-coming more complex, quick-easy human-computer interaction products has become more and more people’s favor and demand. How to get the information in the most natural, conve-nient and efficient way to interact with computers for processing and transmission, has become a problem urgently to be solved, while it is desired to have portable computing devices without a keyboard. Speech recognition is an effective way to address these issues.This thesis deals with speaker-independent isolated-word speech recognition and the main contributions are summarized as follows:1) We introduce the basic theory of speaker-independent isolated-word speech recogni-tion and analyze the basic work processes. Research on algorithms, including speech enhance-ment noise reduction algorithm, voice activity detection algorithm and recognition algorithms is given. In-depth study and analysis of the rationale and derivation of Dynamic Time Warp-ing(DTW) and Hidden Markov Models(HMM) algorithms are carried out, and simulations are performed to compare the advantages and disadvantages of the two algorithms, which demon-strate the the HMM algorithm is more suitable for the system in this project.2) In connection with the inadequate of the traditional double-threshold based on zero-crossing rate of parameters of endpoint detection algorithm, we propose an autocorrelation function-based algorithm, instead of zero-crossing rate parameter of double-threshold speech endpoint detection algorithm, and give the detailed flowchart for it. Matlab simulations show that this proposed algorithm yields a better performance in terms of noise immunity.3) We study the peripherals of Audio Aic23b chip and multi-channel buffer serial periph-erals (McBSP) and how to configure them and accomplish the hardware architecture design based on TI’s TMS320C5509A DSP.4) For isolated word recognition features, we investigate the software design process, work out the flow chart of the various parts of speech recognition, and establish a speech database based on non-ordinary people given command words, in which the training and recog-nition systems are implemented with the command words. A large number of experiments on the platform built, the results demonstrate the feasibil-ity and effectiveness of the system design methods proposed in this thesis.