Voice waveform vector quantization using a competitive algorithm

Abstract

A competitive algorithm is used to train dictionaries for voice waveform vector quantization with a phonetically balanced group of sentences as training sequence. The algorithm follows the standard unsupervised competitive rule used in training neural networks and it is suited to most distortion measures and to any practical dimension. An investigation is carried out to find the best range for the algorithm's parameters and its performance is compared to the results obtained when using the LBG algorithm with the same input data. The testing sequence is another phonetically balanced group of sentences uttered by different speakers.