Continuous speech recognition and syntactic processing in Iranian Farsi language

For the recognition of Iranian Farsi phonemes in continuous speech, a hybrid architecture of neural networks consisting of a self-organizing feature map at the first stage & a multilayer perceptron at the second stage is applied to 26 features extracted from each speech frame: 12 weighted cepstral coefficients, 12 delta-cepstral coefficients, & energy & zero-crossing rate. System architecture & training are described, & a trigram network for text conversion is illustrated. A phoneme recognition rate of 45% was improved to 65% by applying context-dependent correcting production rules. Syntactic processing uses symbolic (artificial intelligence) & connectionist approaches: a context-free recursive descent parser provided eror-free processing, & a four-layer perceptron in auto-associative mode had 80% accuracy in the determination of syntactic errors. 3 Tables, 4 Figures, 51 References. Adapted from the source document