next up previous contents
Next: The labeling and training Up: Task 1.2: Baseline System Previous: Task 1.2: Status

Task 1.2: Technical Description

In this report, we describe the Bref-80 database on which the system is based, the labeling and training procedure and preliminary results we get on phoneme and word recognition experiments. All those experiments were carried out with the context-independent hybrid HMM/MLP technology using the STRUT software [5] for the training and the phonetic alignment and the NOWAY decoder for large vocabulary recognition. BREF-80 [3] is a large read speech corpus from 80 speakers. The text material was selected from the French newspaper Le Monde so as to provide large vocabularies (over 20,000 words) and a wide range of phonetic contexts. As Bref contains 1115 distinct diphones and over 17,500 triphones, it can be efficiently used to train phonetic models. The base lexicon, represented by 35 phonemes, was obtained using a text-to-phoneme tool and was manually verified. The lexicon was extended in order to deal with potential liaisons between words.
The training set used throughout our experiments consists of 3737 sentences (3363 sentences for the training and 374 for the cross-validation) from 56 speakers (approximately 9 hours of speech) and the test set consists of 144 sentences from 8 speakers (4 males, 4 females).



 
next up previous contents
Next: The labeling and training Up: Task 1.2: Baseline System Previous: Task 1.2: Status
Christophe Ris
1998-11-10