| Popis: |
We have proposed and evaluated a novel approach for online speaker adaptation of an acoustic model based on face recognition. Instead of traditionally used audio-based speaker identification we investigated the video modality for the task of speaker detection. A simulated on-line transcription created by a Large-Vocabulary Continuous Speech Recognition (LVCSR) system for online subtitling is evaluated utilizing speaker independent acoustic models, gender dependent models and models of particular speakers. In the experiment, the speaker dependent acoustic models were trained offline, and are switched online based on the decision of a face recognizer, which reducedWord Error Rate (WER) by 12% relatively compared to speaker independent baseline system. |
