Time-frequency coding of the line spectrum frequencies

Abstract

Line spectrum frequencies (LSF) are one of the most popular representations of the short-time spectrum envelope in speech coding applications. LSF of adjacent speech frames are usually highly correlated, so the bit-rate for spectral envelope coding can be reduced by interframe coding. Several coding methods, such as switched-adaptive interframe vector prediction, vector predictive quantization, segment quantization, and matrix quantization exploit the interframe redundancy. A new method of low bit-rate interframe LSF coding based on the discrete wavelet transform is proposed in this paper. Prior to coding, each of the line spectrum frequencies is transformed to wavelet domain by treating it as a time-varying waveform. By choosing the appropriate mother-wavelet, the higher level approximation coefficient of the wavelet decomposition (residual signal) comprises most of the LSF waveform energy. Since these coefficients, which carry most of the LSF information, are time decimated by a factor of 2 to power of decomposition level, significant bit reduction in coding can be obtained. Several wavelets were investigated, identifying the one with the best energy concentration. Spectral distortion caused by neglecting all of the detail coefficients was also calculated as a function of wavelet type and decomposition level.