 |
ITRW on
|
 |
|
ITRW on
|
|
To improve our phone recognizer (38.8% accuracy on Switchboard test data), we investigated the power of MLLR adaptation on HMMs estimated from Switchboard training data. We evaluated its effectiveness on three tasks: speaker adaptation within the Switchboard corpus, environment adaptation for performing recognition on WSJ utterances and language adaptation for performing recognition on BREF utterances (clean read French speech). We used a single iteration of supervised MLLR and a global regression matrix. Two results were surprising: first, although Switchboard speaker adaptation improved accuracy globally (0.69% absolute improvement), it was detrimental for certain speakers; second, environment adaptation was more successful than speaker adaptation in the sense that using WSJ test and adaptation data yielded greater accuracy than using Switchboard test and adaptation data. For language adaptation, a situation in which training and test data were mismatched in environment, speaker, speaking style and language, MLLR could increase the baseline accuracy from 18.4% to 25.4%. Our models were standard HMMs with 5400 4-component Gaussian mixture pdf’s.
Bibliographic reference. Ouellet, Pierre / Dumouchel, Pierre (2001): "Experiments with MLLR applied to switchboard models", In Adaptation-2001, 29-32.