Time-frequency vector quantization with application to isolated world recognition

In low-complexity speaker-independent isolated word recognition systems based on vector quantization (VQ) with multiple codebooks, the performance of the VQ has a big impact on the overall performance of the system. This paper introduces two techniques for combining spectral and temporal information in the VQ process, with the objective of improving the recognition performance, while maintaining or decreasing the storage requirement. The proposed techniques are compared to an existent method based on the probability distribution of the time of occurrence of spectral vectors in the quantization process. The experimental results show that the proposed methods improve significantly the recognition performance and have similar or lower memory requirements to the reference method