소음환경에서의 단독음 인식에 관한 연구 = A Study on Measure of Word Recognition in Noise

다국어 초록 (Multilingual Abstract)

This thesis investigates the noise robustness of four LPC based spectral matching measures applied to the problem of speech recognition.
The spectral matching measures of interest include Log Likelihood Ratio, cepstral distance measure, spectral slope distance measure and weighted cepstral distance measure.
In order to evaluate the recognition performance of the distance measure for noisy speech. speaker independent and speaker dependent isolated word recognition test in vehicle noise environments were carried out. The references are generated using the speech signals recorded in a relatively silent laboratory environments.
The experimental results show that Log Likelihood Ratio is the best for clean speech and spectral slope distance measure is the most robust for the speech signal corrupted by vehicle noise. Weighted cepstral distance measure has shown stable recognition rate for the variation of decision rules and number of reference pattern.