A method and device for speech-to-text conversion. From a given speech, the fundamental tone is extracted. A model of the speech is further created from the speech. In the model, a duration reproduction in words and sentences is obtained. The duration reproduction is compared with a segment duration in the speech. Information which decides which type of accent exists is obtained from the comparison from which a text with sentence accent information is produced. Further, in creating the model lexical information, which may indicate accents in words, can be utilized. Syntax analysis of the model can further by utilized to model a fundamental tone in a sentence. Such operations allow the re-recognition utilizing the accent and stress after a decision as to the type of accent which exists.