DWT-based phonetic groups classification using neural networks
Tuan Van Pham, Gernot Kubin · 2006
This paper presents an improvement of the discrete wavelet transform (DWT)-based phonetic classification algorithm by using neural networks (NN) to learn optimal thresholds for speech classification. Two feedforward NNs (two layers) operate on input features extracted from speech frames (10 ms length) by DWT and statistical measurement in order to classify these frames as transient, voiced vowel, voiced consonant and unvoiced consonant categories. Hard thresholds in our earlier paper are used to detect silence and voiced closure intervals. The new algorithm is tested with the TIMIT database and compared with other algorithms to demonstrate its superior performance.