Speaker-invariant phoneme recognition using multiple neural network models

The authors describe the architecture of a neural network-based ASR (automatic speech recognition) system for extracting speaker-independent features and for recognizing a special class of speech sound, such as the vowel and diphthong sounds. Speaker-invariant morphological properties that are presented in speech spectral patterns are extracted using neural networks. The system considered uses a variation of a neocognitron network model for morphological feature extraction and a perceptron model for feature classification. Some experimental performance results for the proposed system are included