A Very Low Bit-rate Speech Coder Based on GMM
WU Hua-yu · Optoelectronic Technology · 2007
A novel very low bit-rate speech coder based on Gaussian mixture model(GMM),which is used to parameterize the short-time speech spectrum envelope,is proposed in this paper.In the coding procedure,speech signal is firstly pre-emphasized and segmented.Secondly,the segmented speech is transformed to spectrum domain and the spectrum envelope of the segmented speech is obtained.Then the spectrum envelope is parameterized by GMM.So the segmented speech is represented by the means,covariances and mixture weights of GMM.In the decoding procedure,the spectrum envelope of segmented speech is reconstructed with the inverse method of the coding.Then the speech is synthesized based on the reconstructed spectrum envelope,in which the voiced speech is synthesized by sinusoid model and the unvoiced speech is just synthesized by inverse FFT.Since the segmented speech can be represented by very few parameters of GMM,the bit-rate of the coder is very low.The result of the experiment shows that the proposed speech coder presents a good performance.The quality of the synthesized speech is still satisfying when the bit-rate of the coder is reduced to 2.35 kb/s.