- English
- français
Model Adaptation for Sentence Unit Segmentation from Speech
The sentence segmentation task is a classification task that aims at inserting sentence boundaries in a sequence of words. One of the applications of sentence segmentation is to detect the sentence boundaries in the sequence of words that is output by an automatic speech recognition system (ASR). The purpose of correctly finding the sentence boundaries in ASR transcriptions is to make it possible to use further processing tasks, such as automatic summarization, machine translation, and information extraction. Being a classification task, sentence segmentation requires training data. To reduce the labor-intensive labeling task, available labeled data can be used to train the classifier. The high variability of speech among the various speech styles makes it inefficient to use the classifier from one speech style (designated as out-of-domain) to detect sentence boundaries on another speech style (in-domain) and thus, makes it necessary for one classifier to be adapted before it is used on another speech style. In this work, we first justify the need for adapting data among the broadcast news, conversational telephone and meeting speech styles. We then propose methods to adapt sentence segmentation models trained on conversational telephone speech to meeting conversations style. Our results show that using the model adapted from the telephone conversations, instead of the model trained only on meetings conversation style, significantly improves the performance of the sentence segmentation. Moreover, this improvement holds independently from the amount of in-domain data used. In addition, we also study the differences between speech styles, with statistical measures and by examining the performances of various subsets of features. Focusing on broadcast news and meeting speech style, we show that on the meeting speech style, lexical features are more correlated with the sentence boundaries than the prosodic features, whereas it is the contrary on the broadcast news. Furthermore, we observe that prosodic features are more independent from the speech style than lexical features.
- LIDIAP-REPORT-2006-002
Reference
Record created on 2010-02-11, modified on 2017-05-10