Voice identifying system and compression method of characteristic vector set for voice identifying system

Abstract

In the process of obtaining codebook by using the clustering algorithm for the voice eigenvector set, the step of dynamic uniting and splitting subclass is added based on the number of vectors in the set and the total distance measured of vectors in the set. Thus, the sum of the distance measured between the vectors and the corresponding code words in the set is reduced after clustering algorithm to raise accuracy of the clustering algorithm. The method guarantees the recognition performance and reduces the storage quantity of the system. The invention also discloses a voice recognition system, which uses the feature codebook to replace the acoustic model. Gauss probability is not needed to calculate in the decoding process, and the probability value can be obtained by looking up the probability table stored in advance. Thus, the quantity of decoding operation is reduced greatly as well as the recognition speed is raised.

Description

Technical field [0001] The invention relates to a speech recognition system and a compression method of a feature vector set used in the speech recognition system. Background technique [0002] Almost all current speech recognition systems use methods based on statistical pattern recognition. In all speech recognition systems, it is necessary to convert the time-domain sound waves of the speech input into a digitized vector feature to describe and distinguish different pronunciations, which we call Speech feature, based on the feature to build a sound model for all pronunciations, which we usually call an acoustic model in the field of speech recognition. All speech recognition systems must have an acoustic model; at the same time, for large vocabulary continuous speech recognition systems, a language model is also required. The purpose of speech recognition is to give a string of sound feature sequences as input conditions, use acoustic models and language models, and use search...

AI Technical Summary

Problems solved by technology

However, if CDHMM is used in a speech recognition system with a large vocabulary, the Gaussian probability needs to be calculated multiple times when the decoding operation unit performs decoding. Usually, the amount of calculation required in the decoding process is concentrated on the calculation of the Gaussian probability, which requires a lot of Calculations

When large-vocabulary speech recognition is performed on embedded hardware platforms with limited resources such as mobile phones, the response speed of the speech recognition system will be very slow, which cannot meet the needs of actual use.

Images