Reconnaissance de la parole guidée par des transcriptions approchées
Benjamin Lecouteux, Georges Linarès, Pascal Nocera, Jean-François Bonastre.

In many cases, a descriptive transcript can be associated to speech signal : movies subtitles, scenario and theatre, summaries and radio broadcast, rearranged transcription for political debates. Transcripts correspond rarely to the exact word utterance. An actor plays differently each play and a speaker is sometimes far of the promptor. The goal of this work is whether to align the given transcript when it matches the speech signal or to fall back on an automatic speech recognition (ASR) adapted with the transcript language. There are multiple applications : to help deaf people following a play with closed caption aligned to the voice signal (with respect to performers variations), to watch a movie in another language using aligned closed caption, to transcript in real time debates or meetings.