An hybrid network system for continuous speech recognition is presented, based in subword units. At the first stage we detect the order, duration and likelihood of the phoneme sequence with a phoneme spotting system based in Hidden Markov Models. At the second level, a Recurrent Neural Network performs the phoneme decoding by selecting the chain of phonemes that forms a legal word. The network was able to learn the underlying phonetic grammar and could reduce the false alarm rates.
Keywords: Continuous Speech Recognition; Hybrid Network; Phoneme Spotting System; Recurrent Neural Network; Grammar learning.