Multi-layer Perceptron for Isolated Mandarin Vocal Number Recognition
Zhong Hui Lin · Journal of Circuits and Systems · 2000
Multi-layer Perceptron (MLP) for isolated Mandarin vocal number recognition is studied from the perspective of template matching. A new training method is proposed for cases with very limited training samples. Research was conducted on such aspects as overcoming the dynamic characteristic of speech, speech feature extraction, training algorithms and strategies etc. Recognition accuracy (without rejection) of 95.7% for speaker dependent case and of 93.0% for speaker independent case has been achieved.