960 hour English speech data. Book reading speech.
Data prepare
Use one of the options:
-
Prepare data with Kaldi (default in results)
bash local/data_kaldi.sh -h
-
Prepare data with
torchaudio
: run following command to get helpbash local/data.sh -h
Summarize experiments here.
Evaluated by WER (%)
EXPID | dev-clean | dev-other | test-clean | test-other |
---|---|---|---|---|
rnnt + transformer lm | 1.81 | 4.03 | 1.94 | 4.39 |
ctc-crf + transformer lm | 2.05 | 4.54 | 2.25 | 4.73 |