Multi-Level Improvement for a Transcription Generated by Automatic Speech Recognition System for Arabic

In this paper we will propose a novel approach to improving an automatic speech recognition system. The proposed method constructs a search space based on the relations of semantic dependence of the output of a recognition system. Then, it applies syntactic and phonetic filters so as to choose the most probable hypotheses. To achieve this objective, different techniques are deployed, such as the word2vec or the language model Recurrent Neural Networks Language Models (RNNLM) or ever the language model tagged in addition to a phonetic pruning system. The obtained results showed that the proposed approach allowed to improve the accuracy of the system especially for the recognition of mispronounced words and irrelevant words.