ASR Domain Adaptation

In this demo, you'll see several audio files, split into two datasets. You'll be able to compare the predictions of a model trained on the first dataset with the ones from the same model which has been adapted on the second dataset. You can also listen to the audio recordings to see how different these two datasets are.

Explanation of the demo

Acoustic models trained on a certain type of data work really well on similar samples, however, as soon as they encounter different recording conditions, they are lost. Adapting the model allows basically recycling it to work on these new conditions instead of simply throwing it away. Moreover, this approach generally gives better results than a model trained from scratch, because the model will have a meaningful initialization instead of a random one.

In our case, we also use a Domain Adversarial Network to force the adapted model to generalize on the two domains so that it can work as well on one as on the other.

Demo information

December 2018
Guillaume Tâche
Master Student