Speech Recognition Using Kohonen Neural Networks, Dynamic Programming and Multi-Feature Fusion.

摘要

The purpose of this thesis was to develop and evaluate the performance of a three-feature speech recognition system. The three features used were LPC spectrum, formants (F1/F2), and cepstrum. The system uses Kohonen neural networks, dynamic programming, and a rule-based, feature-fusion process which integrates the three input features into one output result. The first half of this research involved evaluating the system in a speaker-dependent atmosphere. For this, the 70 work F-16 cockpit command vocabulary was used and both isolated and connected speech was tested. Results obtained are compared to a two-feature system with the same system configuration. Isolated-speech testing yielded 98.7 percent accuracy. Connected-speech testing yielded 75/0 percent accuracy. The three-feature system performed an average of 1.7 percent better than the two-feature system for isolated-speech. The second half of this research was concerned with the speaker-independent performance of the system. First, cross-speaker testing was performed using an updated 86 word library. In general, this testing yielded less than 50 percent accuracy. Then, testing was performed using averaged templates. This testing yielded an overall average in-template recognition rate of approximately 90 percent and an out-of-template recognition rate of approximately 75 percent.