Dynamic bit allocation in CELP excitation coding

Abstract

The excitation coding, i.e., the LTP (long term predictor) and the innovation coding, requires a large part of the overall bit rate in a CELP (code-excited linear prediction) coder. A method to reduce the excitation coding bit rate is proposed. The fact that the pitch varies only slowly during voiced segments of speech can be exploited to design powerful dynamic bit allocation schemes for the excitation sequence. The bit allocation is determined by two methods. In one method, the LTP index is Huffman-coded. This makes the LTP code book require only a small number of bits during speech segments with stable pitch frequency, i.e., voiced segments. In the other method, a high rate approximation for assigning various numbers of innovation code words for each LTP index is derived. As a complement to dynamic bit allocation, a search method for the LTP index is developed that takes into account the number of innovation code words assigned to each LTP index, in the search for an optimal LTP sequence. Simulations are included that show that with these methods the bit rate can be reduced by 400 bit/s with no changes in speech quality.<>