ON THE TRAINING STRATEGIES OF NEURAL NETWORKS FOR SPEECH RECOGNITION
1992 International Joint Conference on Neural Networks, IJCNN 1992, Maryland, Amerika Birleşik Devletleri, 7 - 11 Haziran 1992, cilt.4, ss.749-754, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Cilt numarası: 4
- Doi Numarası: 10.1109/ijcnn.1992.227228
- Basıldığı Şehir: Maryland
- Basıldığı Ülke: Amerika Birleşik Devletleri
- Sayfa Sayıları: ss.749-754
- Boğaziçi Üniversitesi Adresli: Evet
Özet
Speech signals which are used as training data of neural network are limited and new speakers, new vocabulary, new dialects, and new conditions are continuously encountered. Thus, a speech recognition neural network has to be able to extract the invariant features of speech signal. In this study, we investigate how to introduce invariant features to the networks using conventional back propagation (BP), K-neighbor interpolation training (KNIT) with a number of time-shifted examples (TSE) of the same training sample. The TSE are employed for training of a multilayer perceptron (MLP) and a time-delay neural network (TDNN) structure in order to enrich the training sample set covering larger area of phoneme sample space. Speaker-dependent/b, d, g, m, n, N/ (from 5240 Japanese words) phoneme recognition experiments are performed. It is concluded that the BP training with the TSE improves the performance of a MLP but for a TDNN structure, this is not always the case. The effectiveness of the KNIT algorithm with the TSE depends on the network and the neighboring configuration. The BP and KNIT with TSE algorithms are comparable, but when robustness is the issue, the KNIT becomes advantageous.