Harmonic coding of speech at low bit rates
V. Cuperman, P. Lupini · 1995
Activity in research relating to the compression of digital speech signals has increased markedly in recent years due in part to rising consumer demand for products such as digital cellular telephones, personal communications systems, and multimedia systems. The dominant structure for speech codes at rates above 4 kb/s is Code Excited Linear Prediction (CELP) in which the speech waveform is reproduced as closely as possible. Recently, however, harmonic coding has become increasingly prevalent at rates of 4 kb/s and below. Harmonic coders use a parametric model in an attempt to reproduce the perceptual quality of the speech signal without directly encoding the waveform details. In this thesis, we address some of the challenges of harmonic coding through the development of a new speech codec called Spectral Excitation Coding (SEC). SEC is a harmonic coder which uses a sinusoidal model applied to the excitation signal rather than to the speech signal directly. The same model is used to process both voiced and unvoiced speech through the use of an adaptive algorithm for phase dispersion. Informal listening test results are presented which indicate that the quality of SEC operating at 2.4 kb/s is close to that of existing standard codecs operating at over 4 kb/s. The SEC system incorporates a new technique for vector quantization of the variable dimension harmonic magnitude vector called Non-Square Transform Vector Quantization (NSTVQ). NSTVQ addresses the problem of variable-dimension vector quantization by combining a fixed-dimension vector quantizer with a set of variable-sized non-square transforms. We discuss the factors which influence the choice of transform in NSTVQ, as well as several algorithm features including single-parameter control over the tradeoff between complexity and distortion, simpler uses of vector prediction techniques, inherent embedded coding. Experimental results show that NSTVQ out-performs several existing techniques in terms of providing lower distortion along with lower complexity and storage requirements. Results are presented which indicate that NSTVQ used in the Improved Multiband Excitation (IMBE) environment could achieve equivalent spectral distortion while reducing the overall rate by 1000-1250 bits per second.