Continuous Speech Recognition Algorithm Research And Implementation On Embedded Systems

Automatic Speech Recognition is currently a hot spot among both the academic and the engineering community. The purpose of the research is to make the machine "understanding" of human discourse, hence allowing direct input method to replace indirect input ones (namely, the keyboard, hand-writing pad, scanner and others) in a more natural and more efficient way, further, even to eliminate the indirect input methods. ASR is a cross-disciplinary, with man disciplines involved in. In recent years, academics and engineers from different countries have achieved continuous progress in this area, which brings this important technology from the lab to the consuming market gradually. ASR will enter the industry, home appliances, communications, automotive electronics, medical, family services, consumer electronics and other fields. Particularly, on the embedded systems based core end consumer electronics products, ASR is becoming an important human-machine interface and smart-assistance method.This paper studies most aspects of algorithms and signal processing flows in ASR. First, starting by the study of isolated word recognition algorithm, it discusses the signal flow and algorithm theory including the speech generation mechanism and speech signal acquisition, denoising, transformation, parameter extraction, training and pattern-matching recognition. For improving the recognition accuracy of important mathematical models, the chapter carried out a detailed analysis of HMM; Second, it introduces in the concept of the continuous ASR algorithms and a number of new issues, such as language model (LM), grammar rule, etc. were discussed. And based on the W3C standard specification SRG, the chapter designed a set of experimental rules file syntax specification and hence analyzed of a case grammar file. After the success verification to the above theory algorithms on the GNU Octave platform, a suite of programs have been developed on the Desktop PC to benchmark the performance using C/C++language. Hence, a performance benchmark has been done using a speech database:LetsGoData, provided by the CMU. Hence the analyses of the recognitions result and several improvements on the algorithm have been proposed. Finally, several improvement modifications have been proposed.After the work under the desktop PC has been completed, the porting process from the PC platform to the embedded platform has been described. The performance bottleneck on the embedded platform has been analyzed, too. To the problems, the paper raised a specific recommended method to improve the speed and accuracy performance on the embedded platforms without float point coprocessors. And a fixed-point algorithm has been proposed to replace the standard software float-point emulation library for speed performance improvement. The experiments show out that with some accuracy performance loss, the real-time performance could be improved obviously through some algorithm modification and parameter fine turn.