Speech Recognition System Based On An Improved HMM Algorithm

One of today’s academic research hotspots is the use of biometric technology toidentify the identity of human, while speech recognition is an important branch of artificialintelligence technology. Speech recognition uses the correlation between the characteristicparameters and the human voice, as well as the internal relation of language syntax, toachieve a match with a voice template library. Compared to image recognition and otherartificial intelligence technologies, voice recognition gains a reputation of quickidentification, rapid human-computer interaction, great augmentabil-ity, low hardware dependence, good learning ability, low cost etc., so that variousindustries including safety industry and all other situations requiring authentication canapply speech recognition. With the future development of artificial intelligence technology,voice recognition can be used as an excellent man-machine interactive interface, whichcharacteristic gives it an unlimited development prospect. In order to extract the biologicalcharacteristics, the paper firstly analyzes voice vocal model and auditory model, then use arefined HMM re-estimation algorithm to train the input speeches so as to make up thespeech template, finally use the HMM recognition algorithm to find a match in thetemplate with an unknown speech.This paper firstly studies the vocal mechanism and analyzes the speech component toget the effective voice segments without the silent segment. Then, according to the study inline with the human auditory model, the paper gets a set of parameters capable ofrepresenting the voice characteristics, Mel cepstrum model (Mel). And in contrast withLPC model parameters which is another speech characteristics parameters commonly used,we conclude the advantage of Mel model used in speech recognition.Hidden Markov model is used in the process of building up the speech template aswell as speech recognition. Through analyzing the principle as well as the defects of HMM,this pater puts forward a refined HMM re-estimation algorithm which imports the geneticalgorithm, so that the process of building up the the speech template becomes more robustand accurate.The paper uses HTK toolkit as the base frame, which provides a basic realization ofHMM model. Then we combine it with the genetic algorithm software toolkit GALib tooptimize the speech training process of Baum-Welch algorithm. In spite of the weakness ofrelatively lower convergence speed than the classic HMM re-estimate algorithm, ourrefined HMM training algorithm has a higher recognition accuracy because of excludingthe default that ordinary HMM training algorithm sometimes lead to a locally optimumresult and it’s more easily to find the maximum coefficients.