Automatic Speech Recognition on WSJ and CHiME-4 (test)
50.2WER (-5 dB)Conv-TasNet + FM-Refiner
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Conv-TasNet + FM-RefinerSE front-end=Conv-TasNet, FM-Refiner=true2026.01 | 50.2 | 40.9 | 36.9 | 33 | 28.8 | 26.8 | 36.1 | |
| Conv-TasNetSE front-end=Conv-TasNet, FM-Refiner=false2026.01 | 54.2 | 44.8 | 39.7 | 35.1 | 31 | 27.7 | 38.75 | |
| SGMSE+ + FM-RefinerSE front-end=SGMSE+, FM-Refiner=true2026.01 | 54.2 | 43.7 | 39.1 | 34.4 | 29.9 | 26.5 | 37.97 | |
| DEMUCS + FM-RefinerSE front-end=DEMUCS, FM-Refiner=true2026.01 | 55.9 | 46.3 | 42.3 | 35.8 | 30.2 | 26.7 | 39.53 | |
| SGMSE+SE front-end=SGMSE+, FM-Refiner=false2026.01 | 56.2 | 45.9 | 40.8 | 36.1 | 31.3 | 27 | 39.55 | |
| DEMUCSSE front-end=DEMUCS, FM-Refiner=false2026.01 | 60.9 | 49.2 | 43.7 | 38 | 31.7 | 27.2 | 41.78 | |
| FM-RefinerSE front-end=None, FM-Refiner=true2026.01 | 80.9 | 73.5 | 65.9 | 56.4 | 45.6 | 35.2 | 59.58 | |
| UnprocessedSE front-end=None, FM-Refiner=false2026.01 | 85.6 | 79.7 | 73.4 | 63.9 | 51.7 | 39.3 | 65.6 |