Automatic Speech Recognition on TED-LIUM (test)
2.9WERPhi-4-MM
Evaluation results
| Method | Links | |
|---|---|---|
| Phi-4-MMModel Type=Omni-LLM2026.01 | 2.9 | |
| Gemini-2-FlashModel Type=Omni-LLM2026.01 | 3.01 | |
| Speech-HandsFramework=Qwen2.5-Omni, External Reference=Canary-1b-v22026.01 | 4.21 | |
| Whisper-v2-largeModel Type=ASR model2026.01 | 4.32 | |
| Speech-HandsFramework=Qwen2.5-Omni, External Reference=Parakeet-tdt-0.6b-v32026.01 | 4.37 | |
| Speech-HandsFramework=Qwen2.5-Omni, External Reference=Whisper-v2-large2026.01 | 4.45 | |
| Canary-1b-v2Model Type=ASR model2026.01 | 4.78 | |
| Parakeet-tdt-0.6b-v3Model Type=ASR model2026.01 | 4.9 | |
| ConformerXXL-RNNT-P + Downstream NST (Non-filtered)2021.09 | 5 | |
| ConformerXXL-RNNT-PS#Evaluation Protocol=<unk> token removal2021.09 | 5 | |
| Qwen2.5_omniModel Type=Omni-LLM2026.01 | 5.17 | |
| SOTAEvaluation Protocol=Punctuation removal2021.09 | 5.2 | |
| ConformerXXL-RNNT-P + Downstream NST2021.09 | 5.2 | |
| ConformerXXL-LibriLight2021.09 | 5.7 | |
| GPT-4o-voiceModel Type=Omni-LLM2026.01 | 5.79 | |
| ConformerXXL-RNNT-P2021.09 | 5.9 | |
| Multi-Embedding Subset Selection (Conformer-Large)Subset=5%, Model=Conformer-Large, Approach=MMR, Embedding=Fusion2026.03 | 5.9 | |
| GER (Cascaded)Framework=Qwen2.5-Omni, External Reference=Parakeet-tdt-0.6b-v32026.01 | 6.09 | |
| GER (Cascaded)Framework=Qwen2.5-Omni, External Reference=Whisper-v2-large2026.01 | 6.15 | |
| GER (Cascaded)Framework=Qwen2.5-Omni, External Reference=Canary-1b-v22026.01 | 6.38 | |
| Conformer-Large (Full)Subset=Full, Model=Conformer-Large2026.03 | 6.5 | |
| MMR-based Data SelectionSubset=5%, Model=Conformer-Small, Approach=MMR, Embedding=WavLM2026.03 | 9.2 | |
| Multi-Embedding Subset SelectionSubset=5%, Model=Conformer-Small, Approach=MMR, Embedding=Fusion2026.03 | 9.4 | |
| MMR-based Data SelectionSubset=5%, Model=Conformer-Small, Approach=MMR, Embedding=Speaker2026.03 | 9.6 | |
| MMR-based Data SelectionSubset=5%, Model=Conformer-Small, Approach=MMR, Embedding=SBERT2026.03 | 9.9 | |
| Random (Baseline)Subset=5%, Model=Conformer-Small, Approach=Random (Baseline)2026.03 | 10.7 | |
| Duration (Baseline)Subset=5%, Model=Conformer-Small, Approach=Duration (Baseline)2026.03 | 10.8 | |
| Conformer-Small (Full)Subset=Full, Model=Conformer-Small2026.03 | 11 | |
| Multi-Embedding Subset Selection + Fine-TuneSubset=5%, Model=Conformer-Small, Approach=+ Fine-Tune, Embedding=Fusion2026.03 | 16.3 | |
| Random (Baseline) + Fine-TuneSubset=5%, Model=Conformer-Small, Approach=+ Fine-Tune2026.03 | 17.8 |