Author/Editor     Štrancar, Andrej; Rozman, Robert; Kodek, Dušan M
Title     Razpoznavanje govora z nevronskimi mrežami
Translated title     Automatic speech recognition using neural networks
Type     članek
Source     In: Zajc B, editor. Zbornik 9. elektrotehniške in računalniške konference ERK 2000. Zvezek B. Računalništvo in informatika, umetna inteligenca, robotika, razpoznavanje vzorcev, biomedicinska tehnika, močnostna elektrotehnika, didaktika, študentski članki; 2000 sep 21-23; Portorož. Ljubljana: Slovenska sekcija IEEE,
Publication year     1999
Volume     str. 173-6
Language     slo
Abstract     In this paper we present a neural network based speech recognition system. The system allows several types of neural networks to be used as a class probability estimator. The system is built with two toolkits. Speech parametrisation and Viterbi search are done in CSLU toolkit, building training and frame level testing of neural networks are done using NICO toolkit. The focus of this paper is on using neural networks with delayed connections, TDNNs (Time Delay Neural Networks) and RTDNNs (Recurrent TDNN) in speech recognition systems. Both types of networks are known to be very successful at phoneme recognition. We performed an experiment that compares these types of networks with MLPs (Multi Laver Perceptron), which are used in most systems. in the same recognition system we integrated MLPs, TDNNs and RTDNNs with various number of neurons and evaluated the results on our Slovenian digits corpora. Initially TDNNS were slightly less successful than MLPs, but after the recurrent connections in hidden layer of TDNNs were added the, resulting networks performed significantly better than MLPs.
Descriptors     SPEECH ACOUSTICS
PATTERN RECOGNITION
NEURAL NETWORKS (COMPUTER)
PHONETICS