Segmental phonetic features recognition by means of neural-fuzzy networks and integration in an N-best solutions post-processing

We present investigations on using segmental phonetic features in an N-best solutions post processing of an HMM based ASR system. These phonetic features are extracted by means of neural-fuzzy networks. Specialized neural-fuzzy networks are defined to recognize specific phonetic features (consonant/vowel, voiced/unvoiced, ...). Each of these neural networks furnishes a segmental coefficient (resulting from the output layers) which enables the computation of a segmental post-processing score for the N-best solutions of an HMM based ASR system. This post-processing is based on the computation of segmental score for each solution respectively under the hypotheses of a correct solution and an incorrect solution. Preliminary experiments were conducted on 3 speaker-independent telephone databases. An error rate reduction up to 20% was achieved on the digit corpus.