Automatic Thai audio/transcription segmentation

This paper proposes an automatic algorithm for segmenting audio and transcription streams to be used in constructing a large vocabulary continuous speech recognition (LVCSR) system. In many cases, LVCSR training data are derived from audio materials with available transcriptions such as preach and news articles. In Thai, these resources are usually in the form of long wave files with their corresponding text articles written with no explicit word nor sentence separation, which is crucial for acoustic and language model training in LVCSR. The proposed algorithm segments a large wave file into small utterances using energy detection. The transcription is then aligned to each utterance using dynamic time warping (DTW) combined with a classification and regression tree (CART) confidence measure over a phone basis. An evaluation shows that the DTW alignment procedure still requires an improvement while the CART confidence measure achieves a promising result.