First International Conference on Spoken Language Processing (ICSLP 90)

Kobe, Japan
November 18-22, 1990

Phonetic Features Extraction Using Time-Delay Neural Networks

Frédéric Bimbot, Gerard Chollet, Jean-Pierre Tubach

Telecom Paris - Dept SIGNAL, C.N.R.S. - URA 820, Paris, France

A. Waibel introduced Time-Delay Neural Networks as a specific neural network architecture that is especially well adapted to the "dynamic nature of speech". We propose here to use low-dimensioned TDNNs for discriminating between phonetic features. We give evaluations of the different performances and we comment them. We also compare direct phoneme recognition scores using a sophisticated classical classifier on one hand, and a medium-size TDNN on the other hand. Extra results obtained after having split our corpus into vowels and consonants are also reported. Experiments are conducted on a set of 5270 phonemes extracted from natural continuous speech uttered by 1 male speaker. Nearly all scores on binary phonetic features range between 90 % and 99 %. More complex tasks provide results between 80 % and 90 %.

Full Paper

Bibliographic reference.  Bimbot, Frédéric / Chollet, Gerard / Tubach, Jean-Pierre (1990): "Phonetic features extraction using time-delay neural networks", In ICSLP-1990, 665-668.