Research On Mandarin Connected Digit Speech Recognition

In recent years, speech recognition technology has achieved a breakthrough, starting from the laboratory to the people’s daily lives. Voice dialing, voice access and other means of human-computer voice interaction is being more and more people understand and use speech recognition to show a huge market value and broad application prospects. In this paper,"0" to "9" ten Arabic numerals in Chinese pronunciation, in summarizing the results of previous studies based on the recognition system to improve performance and achieve a number of Chinese continuous speech recognition system.This paper analyzes the development of voice recognition technology status and future trends in speech recognition at this stage revealed the existence of problems and proposed solutions. On this basis, expounded the principles of speech recognition, focusing on the voice signal generated acoustic models. Part of the speech signal endpoint detection, suggestions to improve the MFCC-based similarity (MFCCS) endpoint detection algorithm effectively improve the system’s noise immunity, and gives the simulation results and analysis, experiments show that the algorithm in the low SNR can be a good endpoint detection results. In the speech signal feature extraction part, details the parameters of linear prediction cepstrum (LPCC) and Mel Frequency cepstrum parameters (MFCC) extraction process. Proposed short-term energy and MFCC mix to form a new weighted feature parameters EMFCC, and as a voice training and recognition of the characteristic parameters. Experiments show that, EMFCC better system than the MFCC recognition performance.Then on the Hidden Markov Model (HMM) and its application in speech recognition conducted in-depth study. HMM assessment of the need to address the problem, the decoding problem, and parameter optimization are analyzed in detail, and describes the solution algorithm used to forward and backward, Viterbi algorithm and Baum-Welch revaluation algorithm.Finally, analysis simulation results which were done on the platform of MATLAB, including the speech signal pre-processing, continuous speech endpoint detection, feature extraction, HMM training and recognition.