On the application of vector quantization to speaker independent isolated word recognition
Florina Rogers · Summit (Simon Fraser University) · 1996
The source coding method referred to as vector quantization (VQ) is used in a speechrecognition system to represent an arbitrary speech spectral vector into one of a fixed number of codeboolt symbols with the benefit of significantly reduced computation in the recognition process. I11 low-complexity spealter-independent isolated-word recognition systems with multiple codeboolts, the performance of the VQ has a big impact on the overall perfornlance of the system. This t,hesis studies different ways of combining temporal and spectral characteristics in the VQ process, with the objective of improving t,he recognition, while maintaining or decreasing the storage requirement. Two met,llods of incorpora.ting time information directly into the codebooks are presented and compared to an existent method, based on considering the probability of the time of occurrence of a given spectral vector in the quantization process. The recognition system implemented to evaluate these methods consists of modules which perform signal pre-processing, feature extraction and vector quantization, with a signal-processing front end based on a bank-of-filters model. The experimental results show that the methods proposed reduce significantly the recognition error rate and have similar memory requirements to the reference method. To Dr. Joel G. Rogers, with gratitude for his personal example and his support.