A neural net acoustic phonetic feature extractor based on wavelets

An experimental speech processing system for extracting acoustic phonetic features from speaker independent continuous speech has been built and tested. The system uses wavelet analysis to preprocess the speech, and a two-layer receptive field neural network to recognize the phonetic features. The recognizer was implemented in software on a desktop computer, and DARPA's TIMIT CD-ROM database of continuous speech was used for both training and testing. Preliminary results, using a network trained to recognize voicing and frication, are presented.<
>