Study On The System Of Mandarin Digit Speech On The Basis Of DSP

In contemporary digital times, the research about mandarin digit speech recognition is more and more noticed, since some aspects by digital technique, such as speech intensification, speech transmission, speech recognition, speech synthesis and speech deposit, are the front parts of speech signal processing, of which mandarin digit speech recognition is an important part, and used in many areas. On this background, the author has done some researches.On the basis of the model of speech generation, the speech signals is analyzed in the time domain, in the frequency domain, and especially in the cepstrum. And then, with the theory of pattern recognition, the fundamental theory of speech recognition is discussed in the thesis.This part emphasizes how to design the system, in which TMS320VC549 is used as its core circuit. With the help of TMS320C54X's EVM, the thesis describes the structure of this system, and expounds the process of the system. Instead of PC, 89C51, the singlechip microcomputer, is used to control TMS320VC549. The communication mode is improve, and the ISA bus is replaced by the parallel port when TMS320VC549 communicates with the computer, so the system, which bases on TMS320VC549, can work out of the computer.TMS320VC549 is used as the core circuit, TLE2064 amplifies the signals, TLC320AC02 converts the analog signal to the digital signal (A/D), TMS320VC549 trains and recognizes the speech signals, and the circuit of LCDdisplays the result.As for the feature of mandarin digit speech, the existing arithmetic is cited todesign the software system, and the design process is described in the part. Here, the shore-time ^relative EFP(Energy-Frequency-Product) is used to make the capsheaf of Chinese speech signal, and the short-time relative EFQ(Energy-Frequency-Quotient) is used to separate its syllable and consonant-vowel segment, and it improves the correct rate. The key is that the tone feature is introduced to be the feature parameter. That is to say, the fundamental frequency, the derivative of fundamental frequency and its derivative are used as the feature parameters. The continuous density hidden Markov model(CDHMM) is adopted, Viterbi and Baum-Welch reestimation algorithms is utilized to train and recognize the speech signals. This system is an effective research to promote the mandarin digit speech recognition to business field.