The investigation of using limited precision on a TDNN for consonant recognition

This paper presents a fixed-point arithmetic implementation of consonant recognition in continuous speech with speaker independence. The most widely used neural network learning algorithm, backpropagation (BP), is utilized to train the neural network. The neural network employed here, time delay neural network (TDNN) consists of small sub networks designed to capture the coarticulatory effects of the speech data. The recognition of unvoiced stop consonants; P, T, K, is investigated by using the TIMIT speech data base with 18 speakers of 6 main dialects. Fixed-point simulation results deviate 1-2% from their floating-point counterparts. Overall success rate for the unvoiced stop consonants by using limited precision is between 80 and 90% for the test set