Robust automatic time alignment of orthographic transcriptions with unconstrained speech

A method for automatic time alignment of orthographically transcribed speech using supervised speaker-independent automatic speech recognition based on the orthographic transcription, an online dictionary, and HMM phone models is presented. This method successfully aligns transcriptions with speech in unconstrained 5 to 10 min conversations collected over long-distance telephone lines. It requires minimal manual processing and generally produces correct alignments despite the challenging nature of the data. The robustness and efficiency of the method make it a practical tool for very large speech corpora.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">&gt;</ETX>