ASR Rescoring on WSJ (test)
1.2WERcompositional oracle
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| compositional oracleOracle type=ocp2023.09 | 1.2 | — | — | — | |
| H2T-LoRABackbone=LLaMA-13B, Fine-tuning setting=fine-tuning2023.09 | 2.2 | — | — | — | |
| ClozeGER + Calibration + Post-processingBackbone=LLaMA-2-7b, LoRA=true, Source Speech=false, Calibration=true, Post-processing=true2024.05 | 2.5 | — | — | — | |
| H2T-LoRABackbone=T5-large, Fine-tuning setting=fine-tuning2023.09 | 2.7 | — | — | — | |
| GERBackbone=LLaMA-2-7b, LoRA=true, Source Speech=false2024.05 | 2.8 | — | — | — | |
| ClozeGER + CalibrationBackbone=LLaMA-2-7b, LoRA=true, Source Speech=false, Calibration=true, Post-processing=false2024.05 | 3.3 | — | — | — | |
| ClozeGERBackbone=LLaMA-2-7b, LoRA=true, Source Speech=false, Calibration=false, Post-processing=false2024.05 | 3.7 | — | — | — | |
| H2T-ftBackbone=LLaMA-13B, Fine-tuning setting=fine-tuning2023.09 | 3.8 | — | — | — | |
| H2T-ftBackbone=T5-large, Fine-tuning setting=fine-tuning2023.09 | 4 | — | — | — | |
| n-best oracleOracle type=onb2023.09 | 4.1 | — | — | — | |
| BaselineBackbone=Whisper2024.05 | 4.2 | — | — | — | |
| LMrank2023.09 | 4.3 | — | — | — | |
| Baseline2023.09 | 4.5 | — | — | — | |
| MATEAlignment Loss=Contrastive2023.05 | 6.01 | 7.46 | — | — | |
| MATE-MSEAlignment Loss=MSE2023.05 | 6.1 | 7.65 | — | — | |
| ME-adapterParameter-Efficient Tuning=true2023.05 | 6.43 | 8.2 | — | — | |
| BERT-text2023.05 | 6.46 | 8.2 | — | — | |
| MATE-NAAlignment Loss=None2023.05 | 6.49 | 8.1 | — | — | |
| WavLM-adapterParameter-Efficient Tuning=true2023.05 | 6.52 | 8.05 | — | — | |
| Frozen-MEParameter-Efficient Tuning=true2023.05 | 6.55 | 8.15 | — | — | |
| Multi-modal-GPT22023.05 | 7.55 | 8.2 | — | — | |
| GPT2-text2023.05 | 7.64 | 8.4 | — | — | |
| LAS rescoring2023.05 | 8.01 | 8.55 | — | — | |
| No rescoring2023.05 | 8.16 | 8.75 | — | — | |
| InstructGPTrescoring_pipeline=P2, frozen=true, task-activating prompting (TAP)=true, base_input=First pass (b)2023.09 | 8.72 | — | — | — | |
| Oracledescription=Lower bound baseline2023.09 | 9.78 | — | — | — | |
| InstructGPTrescoring_pipeline=P2, frozen=true, base_input=First pass (b)2023.09 | 9.97 | — | — | — | |
| BLOOMrescoring_pipeline=P2, frozen=true, task-activating prompting (TAP)=true, base_input=First pass (b)2023.09 | 10.82 | — | — | — | |
| 4-gram LMbase_input=First pass (b)2023.09 | 11.21 | — | — | — | |
| OpenLLaMArescoring_pipeline=P2, frozen=true, task-activating prompting (TAP)=true, base_input=First pass (b)2023.09 | 11.53 | — | — | — | |
| First passdescription=Baseline ASR output2023.09 | 11.87 | — | — | — | |
| BLOOMrescoring_pipeline=P2, frozen=true, base_input=First pass (b)2023.09 | 12.59 | — | — | — | |
| OpenLLaMArescoring_pipeline=P2, frozen=true, base_input=First pass (b)2023.09 | 13.32 | — | — | — | |
| GPT-2rescoring_pipeline=P2, frozen=true, task-activating prompting (TAP)=true, base_input=First pass (b)2023.09 | 27.37 | — | — | — | |
| GPT-2rescoring_pipeline=P2, frozen=true, base_input=First pass (b)2023.09 | 29.56 | — | — | — | |
| Oracle2024.05 | — | — | 2.7 | 1 |