Speech Recognition Of Scene-specific Words For Non-specific People

The rapid development of artificial intelligence has promoted the theoretical research and application field of automatic speech recognition.The voice communication between humans and machines has become one of the current development trends.For example,in the scenario of a toll gate in a highway,the use of specific courtesy terms and frequency of use when communicating with past drivers is one of the important criteria for administrators to evaluate their work.This thesis mainly studies the isolated speech recognition algorithm of specific people and non-specific people,and selects 20 polite words exchanged by the toll collectors and past drivers as the isolated vocabulary for the highway toll station,completing a set of non-specific voices.recognition system.The main research work is as follows:(1)Isolated word speech recognition for a specific person.Select 800 voice files in the corpus,and perform an experiment on the voice recognition of the isolated words based on Dynamic Time Warping(DTW)for each speaker.There are three improved methods: when the endpoint is detected,more thresholds are set.When the Mel Frequency Coding Coefficient(MFCC)is extracted as the feature parameter,the dynamic parameter features of the firstorder difference are added.When the template is matched,the DTW algorithm based on the dynamic programming is used.So the system is more correctly identified.The rate reached 94.6%.(2)Isolated word speech recognition for non-specific people.The 2600 speech files of the corpus are divided into five different training sets and test sets.Audio files of training need some pretreatments,including pre-emphasis ? framing ? adding hamming window and endpoint detection.After 24-dimensional feature parameter extraction.The algorithm of Baum-Welch is used to train those feature parameters.HMM reference templates of 20 specific words are gotten.Then after preprocessing audio files of testing and extracting the feature parameters of audio files of testing,we need to match the feature parameters of audio files of testing to the HMM reference template.The recognition results are conducted to obtain.The experimental verification shows that the recognition rate of HMM-based isolated speech recognition system in this corpus is 92.8%.(3)A non-specific person speech recognition system in the scene of highway toll station is built.The human-computer interaction interface is designed,which could not only recognize the speech files of the corpus offline and local voice,but also recognize the real-time online recordings in 4 seconds.