First International Conference on Spoken Language Processing (ICSLP 90)

Kobe, Japan
November 18-22, 1990

A New Method of Consonant Detection and Classification Using Neural Networks

Shigeru Chiba, Kiyoshi Asai

Electrotechnical Laboratory, Tsukuba Ibaraki, Japan

We propose a new method of phoneme detection and classification using neural networks. Neural networks can automatically learn phoneme spotting and classification by this method if hand-labeled speech data are given. In this method, first, the candidate time positions of consonant in continuous speech were extracted using an acoustic cues which indicated the possibility of the consonant location. Then, the phoneme label corresponding to the running spectra around the candidate point was determined based on hand-labeled results for speech wave. These candidate points were divided into several groups depending on acoustic cues in order to reduce the charge of neural network learning. The neural networks were trained by the backpropagation learning procedure as those were able to label continuous speech similar to hand-labeled results when the running spectra around the consonant candidate points were given. The speaker independent consonant recognition experiment showed the recognition rate of about 80 % for the test data set. This result proved the effectiveness of this method.

Full Paper

Bibliographic reference.  Chiba, Shigeru / Asai, Kiyoshi (1990): "A new method of consonant detection and classification using neural networks", In ICSLP-1990, 1065-1068.