Spectral representation of speech based on mel‐generalized cepstral coefficients and its properties

In the mel-generalized cepstral method of analysis, the spectrum model can be varied continuously from the all-pole type to the cepstral type. It is also possible to include human hearing characteristics. This article discusses the spectral representation by mel-generalized cepstral coefficients, aiming at application of the mel-generalized cepstral analysis to speech coding and analysis/synthesis. By the proposed spectral representation parameters, the stability condition for the synthesis filter can be clearly stated, and stability after quantization can easily be guaranteed. The distribution property and the spectral sensitivity of the proposed method are derived, and the quantization and interpolation performance is compared to that of LSP. As a result of subjective evaluation of the synthesized sound, it is shown that the quantization and interpolation performance of the proposed method is better than that of LSP. © 1999 Scripta Technica, Electron Comm Jpn Pt 3, 83(3): 50–59, 2000