Automatic Speech Recognition on kathbath Tamil
27.1WERSMEAR MoE
Evaluation results
| Method | Links | ||
|---|---|---|---|
| SMEAR MoEProjector Parameters=52.98 M2026.01 | 27.1 | 5.5 | |
| Tied ProjectorProjector Parameters=72.64 M2026.01 | 28.4 | 6.1 | |
| Single ProjectorProjector Parameters=18.16 M2026.01 | 28.7 | 6.3 | |
| Dense EnsembleProjector Parameters=72.64 M2026.01 | 28.7 | 6.1 | |
| WSMSSL Model=XLS-R, Version=Ours2026.02 | 28.86 | — | |
| WSMSSL Model=MMS, Version=Ours2026.02 | 28.93 | — | |
| Utterance-Level MoEProjector Parameters=52.98 M2026.01 | 29.4 | 6.4 | |
| MMSSSL Model=MMS, Version=Baseline2026.02 | 29.75 | — | |
| XLS-RSSL Model=XLS-R, Version=Baseline2026.02 | 30.22 | — | |
| Token-Level MoEProjector Parameters=52.98 M2026.01 | 30.3 | 7.3 | |
| WSMSSL Model=mHuBERT, Version=Ours2026.02 | 30.77 | — | |
| mHuBERTSSL Model=mHuBERT, Version=Baseline2026.02 | 31.82 | — | |
| WSMSSL Model=wav2vec 2.0, Version=Ours2026.02 | 32.09 | — | |
| wav2vec 2.0SSL Model=wav2vec 2.0, Version=Baseline2026.02 | 34.66 | — | |
| WSMSSL Model=HuBERT, Version=Ours2026.02 | 35.23 | — | |
| HuBERTSSL Model=HuBERT, Version=Baseline2026.02 | 36.2 | — | |
| WSMSSL Model=data2vec, Version=Ours2026.02 | 36.84 | — | |
| data2vecSSL Model=data2vec, Version=Baseline2026.02 | 36.98 | — | |
| Lang-Specific ProjectorProjector Parameters=72.64 M2026.01 | 45 | 12.6 | |
| Whisper large-v32026.01 | 58.6 | 13 |