Tue Feb 13 20:13:22 CST 20233.7.120.1.0pytorch 1.7.0beam=20，CER tool：https://github.com/yufan-aslp/AliMeeting
| model | Para (M) | Data (hrs) | Eval (CER%) | Test (CER%) |
|---|---|---|---|---|
| MFCCA | 45 | 917 | 16.1 | 17.5 |
beam=10
with separating character (src)
| model | Para (M) | Data (hrs) | Eval_sp (CER%) | Test_sp (CER%) |
|---|---|---|---|---|
| MFCCA | 45 | 917 | 17.1 | 18.6 |
without separating character (src)
| model | Para (M) | Data (hrs) | Eval_nosp (CER%) | Test_nosp (CER%) |
|---|---|---|---|---|
| MFCCA | 45 | 917 | 16.4 | 18.0 |
Considering the differences of the CER calculation tool and decoding beam size, the results of CER are biased (<0.5%).