8th European Conference on Speech Communication and Technology

Geneva, Switzerland
September 1-4, 2003


Nearest-Neighbor Search Algorithms Based on Subcodebook Selection and its Application to Speech Recognition

Jose A.R. Fonollosa

Universitat Politecnica de Catalunya, Spain

Vector quantization (VQ) is a efficient technique for data compression with a minimum distortion. VQ is widely used in applications as speech and image coding, speech recognition, and image retrieval. This paper presents a novel fast nearest-neighbor algorithm and shows its application to speech recognition. The proposed algorithm is based on a fast preselection that reduces the search to a limited number of code vectors. The presented results show that the computational cost of the VQ stage can be significantly reduced without affecting the performance of the speech recognizer.

Full Paper

Bibliographic reference.  Fonollosa, Jose A.R. (2003): "Nearest-neighbor search algorithms based on subcodebook selection and its application to speech recognition", In EUROSPEECH-2003, 2493-2496.