SPEECH RECOGNITION FOR THE TIMIT DATABASE USING NEURAL NETWORKS

Four types of neural networks have previously been established for speech recognition by the authors and tested on a small, 7 speaker, 100 sentence database. This paper describes their application to the TIMIT database. The networks are: a recurrent network phoneme recogniser, a modified Kanerva model morph recogniser, a compositional representation phoneme-toword recogniser and a modified Kanerva model morph-to-word recogniser. The major result is for the recurrent net, giving a phoneme recognition accuracy of 57% from the sz and sz sentences. The Kanerva morph recognizer achieves 66.2% accuracy for a small subset of the sa and sz sentences. The results for the word recognizers are incomplete.