Automatic feature extraction from spectrograms for acoustic-phonetic analysis

Abstract

Proposes a new approach for automatic feature extraction from spectrograms, which is an essential component of acoustic-phonetic analysis in automatic continuous speech recognition. The method comprised four levels: segmentation, pattern classification, feature recognition and labelling, and a post-processor. There were three types of patterns: fuzzy, formant and silence. The extracted features included voice bar, stripes, cut-off and transitions of the first four formants. Some techniques are presented, such as two special distortion functions used in segmentation, and a peak-iterate function to detect the stripes feature. This software has been implemented as part of a speech knowledge interface, which was an expert system for speech analysis for speaker-independent, continuous speech recognition. It has been tested with a set of data chosen from a spectrogram database; the correct detection rate for most features was over 89%, and in some cases was as high as 98%.<>