We setup two challenges: clean challenge and 0-DB challenge. In the clean challenge, the training and test data are all clean, and participants
are required to use the training/test/cv data provided 'officially' by THCHS-20. In the 0-DB challenge, the test data are corrupted by three types of
nois at the SNR of 0db (script is provided to generate the test data at other levels of noise). This is very challenging and more investigation is required. In both
the challenges, the noise signals can be used.
For each condition (clean/noise), two tasks are defined: word ASR and phone ASR. Although the former is more close to practical usage, the latter can test acoustic models in a more focused way.
All individual and research groups are welcome to send their results on these tasks. Simply send an email to firstname.lastname@example.org to inform your results and paper, then
your results will be on the list.
|Institute || Word ASR Task(WER) ||Phone ASR Task (PER) || mystery
||CSLT@THU || 23.18 ||10.01 || DNN, MPE training
If you download the 'standalone' version of THCHS30, the nosiy test data should have been included in wav.tgz
If you download the 'openslr' version (that is suitable for the Kaldi thchs30 recipe), the 0-db test data should be download via the following links:
Baidu:http://pan.baidu.com/s/1hqKwE00, download from "CSLT public data/thchs30-openslr/test0db.tgz"
Mega :https://mega.nz/#F!idRSjL4A!cnCY0R2NjU77Jr0soe9OgQ, download from "CSLT public data/thchs30-openslr/test0db.tgz"
|Institute || White ||Car || Cafeteria || mystery
|CSLT@THU (Word Task, WER) || 65.23 || 25.12 || 53.38 ||DNN, MPE,DAE 
|CSLT@THU (Phone Task, PER) || 40.04 || 11.81 || 32.39 ||DNN, MPE,DAE 
 THCHS-30 : A Free Chinese Speech Corpus