Investigating Image Processing Based Aligner for Large Texts Application for Under-Resourced Languages

摘要

Speech annotation is a costly and time consuming process because it requires high accuracy. Lightly supervised acoustic modeling solves this problem by making use of approximate transcriptions of speech recordings. In under-resourced languages, the speech recordings are not transcribed entirely and the accuracy of the transcription is poor. In this case, it is necessary an additional segmentation step. We propose a segmentation method that uses image processing techniques in order to spot a text island into a larger one. We also investigate on the effect of several tuning parameters on the method's accuracy.