Cascaded likelihood vector compression of linear predictive models
Yair Shoham · The Journal of the Acoustical Society of America · 1988
Compression of linear predictive spectra has applications in acoustics research, particularly, in communication of acoustical information. Cascaded likelihood vector compression (CLVC) is proposed for coding the spectral parameters of linear predictive models at rates of 20 to 26 bits per model. The system is based on representing the LP all-pole model as a cascade of two lower-order models. The partitioning of the LP polynomial is done in the root domain by clustering the roots into two distinct groups. The compressor uses two codebooks to quantize each of the lower-order models. However, the quantization process is done so as to jointly optimize the overall performance. The likelihood ratio distortion measure is used as a performance criterion. Splitting the LP model into two subsystems dramatically reduces the complexity while the efficiency of vector compression is essentially preserved. Experimental results show an average performance of 1.59, 1.41, 1.24, and 1.10 dB of log-spectral distortion at the rates of 20, 22, 24, and 26 bits per model, respectively. These results indicate high-quality model compression particularly for communication applications where the number of bits assigned to the LP model is very limited.