First International Conference on Spoken Language Processing (ICSLP 90)

Kobe, Japan
November 18-22, 1990

Phoneme Recognition by Pairwise Discriminant TDNNs

Jun-Ichi Takami, Shigeki Sagayama

ATR Interpreting Telephony Research Laboratories, Kyoto, Japan

In this paper, a phoneme recognition method using Pairwise Discriminant Time-Delay Neural Networks (PD-TDNNs) is proposed. In conventional approaches to phoneme recognition based on neural networks, it was found that the difference between training data and testing data degrades recognition performance. To overcome this problem, we developed a phoneme recognition method using PD-TDNNs. Each PD-TDNN has the ability to discriminate between two phoneme categories. In this method, phoneme candidates are selected by judging multiple pair discrimination scores, each of which is obtained from the PD-TDNN. We tested this method on a phoneme recognition task for /b,d,g,m,n,N/. Testing on continuous speech using the PD-TDNNs which were trained with the phoneme data in isolated word utterances, we obtained a first candidate recognition rate of 81.6%, and 96.7% for the cumulative recognition rate up to third candidates.

Full Paper

Bibliographic reference.  Takami, Jun-Ichi / Sagayama, Shigeki (1990): "Phoneme recognition by pairwise discriminant TDNNs", In ICSLP-1990, 677-680.