Modification of the speech feature extraction module for the improvement of the system for automatic lectures transcription

This contribution is about experiments with different speech feature extraction methods and strategies where the goal has been to improve the result and the resulting recognition rate of the speech recognizer of an automatic audio speech signal transcription system. The extraction of speech features is based on MFCC (Mel Frequency Cepstral Coefficients) and PLP (Perceptual Linear Prediction), which are normally used in different transcription systems around the world. The speech recognizer with different speech features has been tested on our speech database where audio (or video) recordings from archives of university lectures are stored. The result from our experiments is that we get higher recognition rate if PLP based audio speech features are used.