Research And Implementation On Keyword Recognition Based On DHMM And VQ

Keyword recognition (KWR) is a kind of automatic speech recognition (ASR) technology, and its aim is to detect and confirm a group of special words decided by special occasions from natural voice flow. As the development of ASR technology, KWR technology has extended to communications, automation, human-computer interaction and information retrieval and other areas. Existing KWR system is mainly based on PC, do not fit the requirement of small size and low power for system on a chip (SOC), therefore, KWR system based on integrated circuits (IC) has become the hotspot recently. IC-based speech recognition systems are achieved primarily through the DSP and FPGA. Since there is no independent intellectual property rights of high-performance DSP in China, from the consideration of cost control, FPGA which has many advantages such as low development cost, small size and so on becomes the first choice to develop speech recognition ASIC.Currently, many software algorithms in the existing KWR system are difficult to be implemented by FPGA hardware circuit. On the basis of the basic principle of KWR and mainstream recognition algorithms, this paper has designed the DHMM and VQ-based KWR system which is easy to be achieved by FPGA hardware circuit through research and analysis of discrete hidden Markov model (DHMM) and the introduction of vector quantization (VQ) model.The main research contents of this thesis are recapitulated as follows:(1) This thesis described the basic principle of HMM, and deeply analyzed the forward-backward algorithm, the Viterbi algorithm, the multi-output parameters re-evaluation Baum-Welch algorithm with multi observation sequence.(2) This thesis described the basic principles and common algorithms of speech signal preprocessing, endpoint detection, feature extraction; and deeply discussed and achieved the feature extraction algorithm of MFCC, state machine method endpoint detection based on hardware circuit.(3) This thesis described the basic structure of existing KWR system; analyzed the training algorithm of discrete model DHMM; designed and realized the KWR system based on DHMM, which was easily implemented by FPGA hardware circuit.(4) To ensure the recognition rate and speed of system, on the basis of researching clustering principle, the initial codebook generation and the best codebook design LBG algorithm of VQ, through the introduction of VQ module, the thesis designed and realized the KWR system based on DHMM and VQ.(5) This thesis completed a large number of the model training experiments; simulated and realized the two kinds of designed KWR systems; tested the performance of the systems; analyzed and compared the experimental results statically.