The invention relates to a method for recognising a phonetic sound sequence or a character sequence, e.g. according to the ASCII standard, comprising the following steps: a) the sequence is fed to a neuronal network, b) in said neuronal network, a sequence of characteristics is formed from the phonetic sequence or character sequence, by taking into consideration phonetic and/or lexical stored information and/or based on a character string sequence (blank characters), c) the character sequence thus formed is compared with a characteristic combination (H) that has a defined statement content, said combination being formed from stored lexical and semantic information, based on the characteristic sequence, d) step c) is repeated using new character combinations until, by the reduction of contradictions, a character combination is found that at least largely corresponds with the character sequence, e) the statement content of the character combination with the least number of contradictions is output as the result and/or an action assigned to the statement content is carried out.

Method and device for recognising a phonetic sound sequence or character sequence

In this paper, we address the difficult problem of the learning verification in multilayer neural networks. Finding the activation/inhibition power of each input feature allows us to build synthetic examples and then to find out the minimal recognized patterns as long as to evaluate the robustness of the system. A small illustration upon character recognition clearly shows the interest in bias reduction. A real world application of transients recognition in underwater acoustic helped us to build more efficient features and to significantly improve the generalization rates.

Learning Verification in Multilayer Neural Networks

A speech rescoring system is developed on a set of phonetic hypotheses produced by a bottom-up knowledge-based decoder An original method to automatically compute a fuzzy membership function from top-down acoustic rules statistics is compared with a possibilistic measure To aggregate the fuzzy degrees into a phonetic score, a multilayer neural network is trained on the results of all the rules in order to detect how these rules characterize different phonemes and then in order to give a weight to each rule The rescoring performance of top-down rules for fricatives is discussed on an isolated-word speech database of French with 1000 utterances pronounced by five speakers

R. Quelavoine

Papers

Rescoring under fuzzy measures with a multilayer neural network in a rule-based speech recognition system