ASR rescoring on LibriSpeech clean (test)
4.73WERTranscormer (base, 20K steps)
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Transcormer (base, 20K steps)Domain=Libri, Steps=20K2022.05 | 4.73 | — | |
| T-TADomain=Libri2022.05 | 5.11 | — | |
| MATEAlignment Loss=Contrastive2023.05 | 5.16 | 6.84 | |
| Transcormer (+ BERT Init)Domain=wikibooks, Steps=20K, Initialization=BERT2022.05 | 5.19 | — | |
| MATE-MSEAlignment Loss=MSE2023.05 | 5.22 | 6.95 | |
| RoBERTa (base)Domain=ROBERTa [13]2022.05 | 5.25 | — | |
| Transcormer (base)Domain=wikibooks2022.05 | 5.28 | — | |
| GPT-2 (345M)Domain=openwebtext, Parameters=345M2022.05 | 5.3 | — | |
| BERT (base)Domain=wikibooks2022.05 | 5.41 | — | |
| BERT-textFine-tuning=Full2023.05 | 5.5 | 7.2 | |
| MATE-NAAlignment Loss=None2023.05 | 5.55 | 7.29 | |
| ME-adapterParameter-Efficient Tuning=true2023.05 | 5.56 | 7.43 | |
| Frozen-MEParameter-Efficient Tuning=true2023.05 | 5.57 | 7.34 | |
| WavLM-adapterParameter-Efficient Tuning=true2023.05 | 5.58 | 7.41 | |
| GPT-2 (117M)Domain=openwebtext, Parameters=117M2022.05 | 5.64 | — | |
| ElectricDomain=wikibooks2022.05 | 5.65 | — | |
| Multi-modal-GPT22023.05 | 6.35 | 7.69 | |
| GPT2-textFine-tuning=Full2023.05 | 6.45 | 7.78 | |
| LAS rescoring2023.05 | 6.7 | 7.99 | |
| No rescoring2023.05 | 6.75 | 8.07 | |
| BaselineDomain=N/A2022.05 | 7.26 | — |