Efficient quantization of vocoded speech parameters without degradation
Masanori Morise, Miyashita Genta · 2019
In a statistical parametric speech synthesis (SPSS) system with a vocoder, the dimensions of speech parameters need to be reduced, and many SPSS systems have used companded speech parameters. This paper introduces quantization algorithms for 3 speech parameters: fundamental frequency (fo), spectral envelope, and aperiodicity. In full-band speech (speech with a sampling frequency above 40 kHz), the dimensions of the spectral envelope and the aperiodicity can be reduced to 50 and 5 dimensions based on previous studies. This paper compares the quantization coding without degradation with speech synthesized by the speech parameters without coding. Efficient quantization would be effective for a study that uses graphics processing unit (GPU) computing because recent GPUs support 16-bit floating-point computing. We did two subjective evaluations. The first evaluation determined the appropriate quantization bits in each speech parameter. We obtained the 9 bit values in fo, 13 bit values in the spectral envelope, and 3 bit values in the aperiodicity. The second evaluation verified the effectiveness of our proposed coding. Since a multiple of eight is generally used for data chunks, we employed the 16 quantization bits for fo, 16 for the spectral envelope, and 8 for aperiodicity in the evaluation. The results showed that our proposed algorithm achieved almost all the same sound quality as the speech parameters without coding.