Frequency domain methods for coding the linear predictive residual of speech signals

Abstract

The most frequently used speech coding paradigm is ACELP, famous because it encodes speech with high quality, while consuming a small bandwidth. ACELP performs linear prediction filtering in order to eliminate the effect of the spectral envelope from the signal. The noise-like excitation is then encoded using algebraic codebooks. The search of this codebook, however, can not be performed optimally with conventional encoders due to the correlation between their samples. Because of this, more complex algorithms are required in order to maintain the quality. Four different transformation algorithms have been implemented (DCT, DFT, Eigenvalue decomposition and Vandermonde decomposition) in order to decorrelate the samples of the innovative excitation in ACELP. These transformations have been integrated in the ACELP of the EVS codec. The transformed innovative excitation is coded using the envelope based arithmetic coder. Objective and subjective tests have been carried out to evaluate the quality of the encoding, the degree of decorrelation achieved by the transformations and the computational complexity of the algorithms.