Correction centralisée des distorsions spectrales de la parole sur les réseaux téléphoniques
Gaël Mahé, Jean-Marc BOUCHER · OpenGrey (Institut de l'Information Scientifique et Technique) · 2002
The aim of this thesis is to compensate for spectral distortions of voice on telephone networks, particularly on the analog parts of the terrestrial network. These distortions are generated by the non-flat transfer functions of the sending and receiving terminals and of the corresponding analog lines. Our purpose is to restore a "timbre" as close as possible to the original voice of the speaker, using a blind equalizer centralized in the network. We propose a spectral equalization algorithm, which consists in matching the long-term spectrum of the processed signal to a reference spectrum (spectrum of the ITU-T Recommendation P.50) in a limited frequency bandwidth (200-3150 Hz). Subjective evaluations show a satisfying restoration of the timbre of the speakers, within the limits of the chosen equalization band. The A-law quantization of the output samples of the equalizer induces however a disturbing noise at the reception end. Two methods are proposed to mask this noise, using a perceptual spectral shaping. The first one is based on the feed-back of the filtered quantization error to the input of the quantizer. The second one explores the temporal sequences of the possible quantization levels, in order to maximize a probabilistic criterion of noise masking, using a Viterbi-like algorithm. A subjective evaluation finally shows on the one hand that the non-reshaped noise is preferred to the reshaped noise, on the other hand that voices with a corrected timbre, even with quantization noise, are preferred to the same voices at the output of a telephone link without timbre correction (and without noise). In order to make the reference spectrum more appropriate to the various speakers' voices, we define two or four classes of speakers, based on their long-term spectra. Classification criteria robust to telephone link distortions are defined. This classification allows using one reference spectrum for each class, instead of the same reference spectrum for the whole population. This leads to a decrease of the spectral distortion induced by the equalizer and, as a consequence, to a significant improvement of the timbre correction for a part of the speakers, in a perceptual point of view.