Fully Recurrent Neural Networks for Phoneme Based Speech Recognition

. Here we report about investigations concerning the application of Fully Recurrent Neural Networks (FRNN) for phoneme based speaker-independent automatic speech recognition. Commonly, artificial neural networks ar used for estimation of phoneme or segment probabilities and compensation of time duration is done with conventional methods of dynamic programming leading to hybrid systems. In this contribution FRNN are proposed for extracting contextual information as well as for compensating time durations. It can be shown that FRNN are powerful and tractable tools to estimate phoneme probabilities and to replace conventional time alignment modules based on viterbi algorithm. Therefore, FRNN leads to a monolithic construction of automatic speech recognizers is so far a promissing conception for hardware implementation. 1. Introduction Automatic speech recognizers (ASR) are faced with two basic problems. The extraction of contextual information between feature vectors and the compensation..