An Interactive Way to Acquire Internet Documents for Language Model Adaptation of Speech Recognition Systems
In this paper, a new methodology for speech corpora definition from internet documents is described, in order to record a large speech database, dedicated to the training and testing of acoustic models for speech recognition. In the first section, the Web robot which is in charge of collecting Web pages from Internet is presented, then the web text to French sentences filtering mechanism is explained. Some information about the corpus organization (90% for training and 10% for test) is given. In the third section, the phoneme distribution of the corpus is presented and comparison is made with others French language studies. Finally tools and planning for recording the speech database with more than one hundred speakers are described.