Speech Compression for Noise-Corrupted Thai Dialects

Abstract

Problem statement: Dialects of Thai are quite different in the speaking styles. Environmental noises play an important role in corrupting the communication speech quality. Two factors affect the coded speech in the present speech communication. It is necessary to investigate how the two factors influence on the speech compression. Approach: In this study, the Multi-Pulse based Code Excited Linear Predictive (MP-CELP) coder and the Conjugate Structure Algebraic Code Excited Linear Predictive (CS-ACELP) coder are selected as the coding methods. This study shows the effects of the six kinds of noise to speech coding quality. The comparison of speech quality of the four coded Thai dialects is conducted. The speech material includes a hundred male speech utterances and a hundred female speech utterances. Four speaking styles include Thai Northern, North Eastern, Southern and Central dialects. Five sentences of Thai speech are chosen. Six types of noise include train, factory, motorcycle, air conditioner, men speaker and women speaker. Moreover, five levels of each type of noise are varied from 0-20 dB. The subjective test of mean opinion score are exploited in the evaluation process. Results: The experimental results show that CS-ACELP gives better speech quality than that of MP-CELP at all three bitrates of 6000, 8600 and 12600 bps. When considering the levels of noise, the 20-dB noise gives the best speech quality, while 0-dB noise gives the worst speech quality. When considering the speech gender, male speech gives better results than that of female speech. When considering the types of dialect, the central dialect gives the best speech quality, while the North dialect gives the worst speech quality. Finally, when considering the types of noise, the air-conditioner noise gives the best speech quality, while the train noise gives the worst speech quality. Conclusion: From the study, it can be seen that coding method, type of noise, level of noise, speech gender and dialect influence on the coding speech quality.