Third International Conference on Spoken Language Processing (ICSLP 94)
It is well-known the limited quality of the reconstructed speech by Code-Excited Linear Prediction (CELP) speech coder at low bit-rate (2400-8000 bps). The major source of audible distortion has been attributed to an inaccurate degree of periodicity of the voiced speech signal. In the present paper we alleviate this drawback by modifying the MSE distance measure, which is not able to capture the periodicity adequately. The new error criterion proposed is a projection similarity measure, which computes a projection distance of original onto coded perceptually weighted voice on a point to point basis. The improvement of the quality of the speech reconstructed by the CELP with projection distance measure has been checked by subjective A-B test. This result emphasizes the perceptual importance of an adequate description of the pitch-pulse waveform of the original (uncoded) LP residue by the CELP-coded one.
Bibliographic reference. Ferrer-Ballester, Miguel A. / Figueiras-Vidal, Anibal R. (1994): "Improving CELP voice quality by projection similarity measure", In ICSLP-1994, 2063-2066.