Transcription automatique de la parole spontanée

Abstract : Automatic speech recognition (ASR) systems currently reach enough performance to be integrated in various applications (human-machine dialogue, information retrieval, automatic indexing ...). However, in the context of large vocabulary speech recognition, which is used eg for transcribing radio broadcast, the quality of transcripts varies depending on the type of speech contained in the documents. Indeed, the ASR system performance are much better when transcribing prepared speech, close to a read text . Transcribing pontaneous speech is much more dificult, as it is characterized by many features (disfluencies, ungrammaticality, decreased the fluidity of speech...). This thesis work is the treatment of spontaneous speech and is part of the EPAC project. The main objective is to propose solutions to improve the ASR performance on this type of speech. We chose to address in our work, spontaneous speech as a particular object of study requiring specific treatments. Thus, in a first step, we propose a tool for automatic detection of spontaneous speech, based on its specificities. This tool is very important because it allows us, in a second time, to propose an approach for acoustic and language model adaptation of the ASR system on spontaneous speech without adding data, by automatically selecting the segments containing this type of speech. The transcript resulting from this adaptation offers recognition hypotheses different from those provided by the base system. The combination of these two proposals transcription show a significant reduction of the word error rate. This need for specific solutions finally facing some of our work toward correcting a specific problem, especially present in French: homophony. We then seek to correct the transcripts provided by an ASR system, using a method offering specific solutions to specific problems of homophony. The approach focuses on correcting errors, to which a particular solution is proposed. This post-processing method of ASR systems corrects some classes of words and homophones, regardless of the ASR system used.
Document type :
Theses
Computer Science. Université du Maine, 2010. French


https://tel.archives-ouvertes.fr/tel-00595465
Contributor : Richard Dufour <>
Submitted on : Tuesday, May 24, 2011 - 6:18:59 PM
Last modification on : Wednesday, May 25, 2011 - 9:03:26 AM

Identifiers

  • HAL Id : tel-00595465, version 1

Collections

Citation

Richard Dufour. Transcription automatique de la parole spontanée. Computer Science. Université du Maine, 2010. French. <tel-00595465>

Export

Share

Metrics

Consultation de
la notice

268

Téléchargement du document

294