ABSTRACT
Speech recognition technology involves phonetics, artificial intelligence, digital
signal processing technology, physiology, mathematical statistics, and other
interdisciplinary knowledge and it is the most convenient way of man-machine
communication. Recently, it is a major research focus that applying the speech
recognition to mobile devices, smart home and other areas. The problem we are
addressing in this paper is that of recognizing isolated word speech on DSP.
Firstly, we create the template library of speech automatically based on the
Cambridge HTK tool sets which is mainly used for speech recognition. The template
library has a simple training process and is free of expansion. Besides, the recognition
rate has improved as we can approach the optimal HMM parameters.
Then we study sampling, digitization, pre-emphasis processing algorithm,
feature extraction algorithm, endpoint detection algorithm, speech recognition
algorithm of the voice signal. To cater for the recognition rate will decline as we
cannot identify the beginning and ending points of the speech signal very well in a
noisy environment, an endpoint detection algorithm combined by automatically
adjusting thresholds algorithm and enhancing the zero-crossing rate algorithm is
utilized, and it is shown to perform with good precision on finding the start and end
points of voice signal, which is simulated by MATLAB. To enable high computing
speed and good real-time to be performed, we use the improved Mel cepstral
coefficients for characteristic parameters. Finally we have realized the HMM-based
non-specific speech recognition system successfully.
Lastly, we realize the non-specific isolated word speech recognition system
based on CCS which is the software development tool of DSP and
ICETEK-DM642-PCI platform which consists of TMS320DM642 processor core and
TLV320AIC23 audio codec.
The experiment results show that the recognition rate is up to 95% in this system.
KEY WORDS：Speech recognition, DSP, Endpoint detection, Mel Cepstrum parameters,
HMM
II
剩余57页未读，继续阅读
资源评论
