Robust tree coding combined with harmonic scaling of speech

Abstract

Efficient speech coders using tree coding combined with harmonic scaling are designed at the rates of 6.4 kilobits/sec (kbps) and 4.8 kbps. A time domain harmonic scaling algorithm (TDHS) is used to compress input speech by a factor of two. This process allows the tree coder to have 2 bits/sample for 6.4 kbps and 1.5 bits/sample for 4.8 kbps in the case of a 6.4 kHz sampling rate. Various window functions and pitch extraction algorithms are compared in the TDHS algorithm. In the backward adaptive tree coder, there are three components of the code generator, including a hybrid adaptive quantizer, a short-term predictor and a pitch predictor. The short-term predictor with a pole-zero structure improves the performance of the tree coder in the error free channel over an all-pole structure. The robustness of the tree coder is achieved by carefully choosing the input of the short-term predictor adaptation. The residual signal shaped by an all-zero filter is used as the input for adaptation. The all-zero filter is obtained by truncating the infinite impulse response of the pole-zero synthesis filter. Another method to improve the robustness of the tree coder is the inclusion of a smoother in the pitch predictor. The coefficients of the smoother with three taps are computed from the estimated autocorrelation function of the pitch synthesizer output. The inclusion of a smoother provides a small improvement in the error free channel and robustness to transmission errors. Subjectively, tree coding combined with TDHS achieves very good quality speech at the rate of 6.4 kbps and good quality speech at 4.8 kbps. The performance of the coding system is evaluated for real speech by means of signal-to-noise ratio (SNR), segmental SNR, frequency domain SNR, spectrograms, and informal listening tests.