8th European Conference on Speech Communication and Technology

Geneva, Switzerland
September 1-4, 2003


Quality Enhancement of CELP Coded Speech by Using an MFCC Based Gaussian Mixture Model

D.G. Raza, C.F. Chan

City University of Hong Kong, China

At low bit rates CELP coders present certain artifacts generally known as hoarse and muffing characteristics. An enhancement system is developed to lessen the effects of these artifacts in CELP coded speech. In enhancement system, the high frequency components (4kHz-8kHz) are reinserted to reduce the muffing characteristics. This is achieved by using an MFCC based Gaussian Mixture Model. The hoarse characteristics are reduced by re-synthesizing the CELP reproduced speech with harmonic plus noise model. The pair-wise listening experiment results show that the re-synthesized wideband speech is preferred over the CELP coded speech. The enhanced speech is affirmed to be pleasant to listen and exhibits the naturalness of the original wideband speech.

Full Paper

Bibliographic reference.  Raza, D.G. / Chan, C.F. (2003): "Quality enhancement of CELP coded speech by using an MFCC based Gaussian mixture model", In EUROSPEECH-2003, 541-544.