First International Conference on Spoken Language Processing (ICSLP 90)

Kobe, Japan
November 18-22, 1990

Formant Extraction Model by Neural Networks and Auditory Model Based on Signal Processing Theory

Kazuaki Obara, Hideyuki Takagi

Central Research Laboratories, Matsushita Electric Industrial Co., Ltd., Osaka, Japan

This paper discribes a formant extraction model constructed by an auditory model for preprocessing and a neural networks (NN) that extracts formant. This auditory model consists of the first stage based on physiological foundings and the second stage based on a hypothetic function. The first stage is consisted of a basilar membrane model of 190ch critical band pass filter bank (CBF), hair cell model of half-wave (HW) rectification and satulating property, and synapse or axon model of low pass (LP) characteristics. The second stage has hypothetical function that has homomorphic processing. Finally, a NN for formant extraction and its experiment are described. This NN extract formant from tge output of auditory model. The resuls displayed in this paper shows that our formant extraction model performs well for not only supervised data but also unsupervised data.

Full Paper

Bibliographic reference.  Obara, Kazuaki / Takagi, Hideyuki (1990): "Formant extraction model by neural networks and auditory model based on signal processing theory", In ICSLP-1990, 1045-1048.