ISCA Archive - An integration of knowledge and neural networks toward a phoneme typewriter without a language model (original) (raw)

An integration of knowledge and neural networks toward a phoneme typewriter without a language model

Yasuhiro Komori, Kaichiro Hatazaki

This paper proposes a phoneme recognizer without any language model. The system is realized as an integration of spectrogram reading knowledge and Time-Delay Neural Networks. The system mainly consists of two parts: a consonant recognition part and a vowel recognition part, in which a sophisticated integration of knowledge and TDNN, is proposed. The knowledge part is mainly used for verification of ciitegories and boundaries. An experiment of speaker-dependent phoneme recognition without any language model, using 2,620 words, showed a 91. 4% recognition rate, a 3. 6% deletion error rate, a 5. 0% substitution error rate and a 20. 7% insertion error rate, for all Japanese phonemes. Keywords: Speech recognition; Spectrogram reading knowledge; Time-Delay Neural Networks; Expert system