Created by: lifelongeek

Librispeech provides fairly large amounts of speech (960hour) which is recorded from various recording environment. So I expect training acoustic model on this corpus will also works for speech in another domain, such as CHiME-3.

For acoustic model, deepspeech model is trained with librispeech corpus. For language model, 4-gram (from http://www.openslr.org/11/) is used.

It achieve 7.6% WER on test-clean, 24.8% WER on test-other.

However, this model is not working to recognize CHiME-3 data.

Clean evaluation dataset got WER 64.2%, Cafe-noise evaluation dataset got 83.7% WER.

What could be the possible reason for degradation?