Efficient Vector Quantization of LSF Parameters

E.V. Krishna Rao, P.G. Krishna Mohan, P.V. Subbaiah, and N.V.N. Prathyusha (India)


Linear Predictive Coding, Line Spectral Frequencies, Vector Quantization, Speech Coding, Spectral Distortion.


This paper describes a new method of efficient Vector Quantization of Line Spectral Frequency (LSF) Parameters for Speech Coding. This method is much better than the existing algorithms like LBG (LINDE, BUZO, GRAY) algorithm. The Speech signal is sampled at 8 KHz and converted into number of frames. Linear Predictive Coding (LPC) Parameters are calculated every frame. These LPC parameters are converted into LSF parameters. Taking the training samples from N files, a long rawcodebook RN is prepared. The hit-books are prepared from M different speech files, the sum of hits of M files hitsum of one rawcodebook is obtained. Repeating the same for N rawcodebooks, appending all these N hitsums create appended hitbook HN. By removing redundant vectors in the appended rawcodebook and its corresponding hits, the Modified rawcodebook (MRN) and Modified hitbook (MHN) are prepared. Then optimum codebook is prepared by rearranging the elements in the Modified rawcodebook (MRN) in the descending order of hits. The LSF Parameters are quantized using optimum codebook and then the index is transmitted to the receiver. Further by removing the unvoiced frames, the LSF parameters of the voiced frames are quantized. The index of the removed unvoiced frames is maintained and the same is introduced at synthesis to reconstruct the original signal. The number of bits/frame vs. Spectral Distortion (SD) is presented with removing unvoiced frames. It can be seen that average SD is very much less in the proposed algorithm

Important Links:

Go Back