Automated Speech Recognition on TED-LIUM V3
3.11WERSeed-ASR (ML)
Evaluation results
| Method | Links | |
|---|---|---|
| Seed-ASR (ML)Language=EN2024.07 | 3.11 | |
| FC-XXLDecoder=CTC, Train Dataset=ASR Set ++2023.05 | 3.54 | |
| Whisper (real label)Adaptation strategy=Finetuning, Supervision type=Supervised2024.05 | 3.6 | |
| FC-XLDecoder=CTC, Train Dataset=ASR Set ++2023.05 | 3.78 | |
| FC-XLDecoder=RNNT, Train Dataset=ASR Set ++2023.05 | 3.86 | |
| FC-XXLDecoder=RNNT, Train Dataset=ASR Set ++2023.05 | 3.92 | |
| Whisper Large-v2Language=EN2024.07 | 4 | |
| STARAdaptation strategy=STAR, Supervision type=Unsupervised2024.05 | 4.1 | |
| TOKreweight (A_t)Adaptation strategy=Self-training, Token-level re-weighting=A_t, Supervision type=Unsupervised2024.05 | 4.3 | |
| FC-XLDecoder=RNNT, Train Dataset=ASR Set2023.05 | 4.49 | |
| FC-XXLDecoder=CTC, Train Dataset=ASR Set2023.05 | 4.64 | |
| UTTfilterAdaptation strategy=Self-training, Filtering=Utterance-level, Supervision type=Unsupervised2024.05 | 4.7 | |
| FC-XLDecoder=CTC, Train Dataset=ASR Set2023.05 | 4.71 | |
| FC-XXLDecoder=RNNT, Train Dataset=ASR Set2023.05 | 4.74 | |
| TOKreweight (G_t)Adaptation strategy=Self-training, Token-level re-weighting=G_t, Supervision type=Unsupervised2024.05 | 4.8 | |
| Whisper (self-train.)Adaptation strategy=Self-training, Supervision type=Unsupervised2024.05 | 4.9 | |
| hyp / Whisper + DPON=4, Con_inf=hyp, Con_train=Whisper, Dropout rate=02026.03 | 4.93 | |
| hyp / Whisper + DPON=5, Con_inf=hyp, Con_train=Whisper, Dropout rate=0.52026.03 | 4.96 | |
| hyp / Whisper + DPON=2, Con_inf=hyp, Con_train=Whisper, Dropout rate=02026.03 | 5.07 | |
| hyp / Whisper + DPON=2, Con_inf=hyp, Con_train=Whisper, Dropout rate=0.52026.03 | 5.17 | |
| hyp / Whisper + DPON=3, Con_inf=hyp, Con_train=Whisper, Dropout rate=0.52026.03 | 5.18 | |
| Whisper (frozen)Adaptation strategy=None, Supervision type=Zero-shot2024.05 | 5.2 | |
| hyp / Whisper + DPON=1, Con_inf=hyp, Con_train=Whisper, Dropout rate=0.52026.03 | 5.32 | |
| hyp / Whisper + DPON=5, Con_inf=hyp, Con_train=Whisper, Dropout rate=02026.03 | 5.34 | |
| hyp / WhisperN=2, Con_inf=hyp, Con_train=Whisper, Dropout rate=0.52026.03 | 5.47 | |
| hyp / GTN=2, Con_inf=hyp, Con_train=GT, Dropout rate=0.52026.03 | 5.59 | |
| GT / GTN=1, Con_inf=GT, Con_train=GT, Dropout rate=02026.03 | 5.6 | |
| hyp / WhisperN=1, Con_inf=hyp, Con_train=Whisper, Dropout rate=02026.03 | 5.62 | |
| GT / GTN=2, Con_inf=GT, Con_train=GT, Dropout rate=0.52026.03 | 5.66 | |
| hyp / Whisper + DPON=1, Con_inf=hyp, Con_train=Whisper, Dropout rate=02026.03 | 5.69 | |
| hyp / Whisper + DPON=4, Con_inf=hyp, Con_train=Whisper, Dropout rate=0.52026.03 | 5.69 | |
| hyp / Whisper + SFT2N=1, Con_inf=hyp, Con_train=Whisper, Dropout rate=02026.03 | 5.76 | |
| hyp / GTN=1, Con_inf=hyp, Con_train=GT, Dropout rate=02026.03 | 5.85 | |
| hyp / Whisper + DPON=3, Con_inf=hyp, Con_train=Whisper, Dropout rate=02026.03 | 5.98 | |
| hyp / Whisper + SFT2N=2, Con_inf=hyp, Con_train=Whisper, Dropout rate=0.52026.03 | 6.1 | |
| GT / GTN=2, Con_inf=GT, Con_train=GT, Dropout rate=02026.03 | 6.73 | |
| hyp / GTN=2, Con_inf=hyp, Con_train=GT, Dropout rate=02026.03 | 6.89 | |
| hyp / Whisper + SFT2N=2, Con_inf=hyp, Con_train=Whisper, Dropout rate=02026.03 | 6.9 | |
| hyp / GTN=3, Con_inf=hyp, Con_train=GT, Dropout rate=02026.03 | 7.05 | |
| hyp / WhisperN=1, Con_inf=hyp, Con_train=Whisper, Dropout rate=0.52026.03 | 7.21 | |
| hyp / Whisper + SFT2N=1, Con_inf=hyp, Con_train=Whisper, Dropout rate=0.52026.03 | 7.26 | |
| GT / GTN=3, Con_inf=GT, Con_train=GT, Dropout rate=02026.03 | 7.35 | |
| Whisper Large-v3Language=EN2024.07 | 7.4 | |
| hyp / GTN=1, Con_inf=hyp, Con_train=GT, Dropout rate=0.52026.03 | 7.47 | |
| Universal-1Language=EN2024.07 | 7.5 | |
| hyp / GTN=4, Con_inf=hyp, Con_train=GT, Dropout rate=02026.03 | 7.74 | |
| hyp / WhisperN=4, Con_inf=hyp, Con_train=Whisper, Dropout rate=0.52026.03 | 7.81 | |
| hyp / WhisperN=3, Con_inf=hyp, Con_train=Whisper, Dropout rate=0.52026.03 | 7.87 | |
| - / -N=0, Con_inf=-, Con_train=-, Dropout rate=02026.03 | 7.89 | |
| GT / GTN=1, Con_inf=GT, Con_train=GT, Dropout rate=0.52026.03 | 7.89 | |
| hyp / Whisper + SFT2N=3, Con_inf=hyp, Con_train=Whisper, Dropout rate=0.52026.03 | 8.01 | |
| hyp / WhisperN=2, Con_inf=hyp, Con_train=Whisper, Dropout rate=02026.03 | 8.15 | |
| hyp / GTN=5, Con_inf=hyp, Con_train=GT, Dropout rate=0.52026.03 | 8.19 | |
| hyp / WhisperN=5, Con_inf=hyp, Con_train=Whisper, Dropout rate=0.52026.03 | 8.5 | |
| hyp / Whisper + SFT2N=5, Con_inf=hyp, Con_train=Whisper, Dropout rate=0.52026.03 | 8.51 | |
| GT / GTN=4, Con_inf=GT, Con_train=GT, Dropout rate=02026.03 | 8.54 | |
| GT / GTN=5, Con_inf=GT, Con_train=GT, Dropout rate=02026.03 | 8.72 | |
| hyp / Whisper + SFT2N=4, Con_inf=hyp, Con_train=Whisper, Dropout rate=0.52026.03 | 9.16 | |
| GT / GTN=4, Con_inf=GT, Con_train=GT, Dropout rate=0.52026.03 | 9.22 | |
| hyp / Whisper + SFT2N=3, Con_inf=hyp, Con_train=Whisper, Dropout rate=02026.03 | 9.3 | |
| GT / GTN=5, Con_inf=GT, Con_train=GT, Dropout rate=0.52026.03 | 9.57 | |
| hyp / WhisperN=3, Con_inf=hyp, Con_train=Whisper, Dropout rate=02026.03 | 10.06 | |
| hyp / GTN=5, Con_inf=hyp, Con_train=GT, Dropout rate=02026.03 | 10.34 | |
| GT / GTN=3, Con_inf=GT, Con_train=GT, Dropout rate=0.52026.03 | 10.42 | |
| hyp / GTN=4, Con_inf=hyp, Con_train=GT, Dropout rate=0.52026.03 | 10.87 | |
| hyp / GTN=3, Con_inf=hyp, Con_train=GT, Dropout rate=0.52026.03 | 12.62 | |
| ChunkFormercontext size [latt, c, r]=[128, 64, 128]2025.02 | 21.32 | |
| ChunkFormercontext size [latt, c, r]=[128, 256, 128]2025.02 | 21.38 | |
| ChunkFormercontext size [latt, c, r]=[256, 128, 128]2025.02 | 21.41 | |
| ChunkFormerdecoding_mode=full-context2025.02 | 23.37 | |
| Efficient Conformerdecoding_mode=full-context2025.02 | 23.4 | |
| Squeezeformerdecoding_mode=full-context2025.02 | 23.5 | |
| Conformerdecoding_mode=full-context2025.02 | 24.03 | |
| hyp / Whisper + SFT2N=5, Con_inf=hyp, Con_train=Whisper, Dropout rate=02026.03 | 72.55 | |
| hyp / WhisperN=4, Con_inf=hyp, Con_train=Whisper, Dropout rate=02026.03 | 87.37 | |
| hyp / Whisper + SFT2N=4, Con_inf=hyp, Con_train=Whisper, Dropout rate=02026.03 | 113.95 | |
| hyp / WhisperN=5, Con_inf=hyp, Con_train=Whisper, Dropout rate=02026.03 | 135.57 |