Automatic Speech Recognition on English DementiaBank Pitt (Eval)
19.25WER (Paraphrase)Enc&Dec Prompts
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Enc&Dec PromptsModel=Whisper (LoRA)2026.06 | 19.25 | 11.76 | |
| Enc Only PromptsModel=Whisper (LoRA)2026.06 | 19.53 | 12.43 | |
| Audio-Textual PromptsModel=Whisper (LoRA)2026.06 | 19.65 | 11.21 | |
| Audio-Only PromptsModel=Whisper (LoRA)2026.06 | 19.73 | 11.32 | |
| Whisper (LoRA)Model=Whisper (LoRA)2026.06 | 20.68 | 12.65 | |
| Supervised (100%)Finetune Strategy=100% Sup.2026.06 | 20.68 | 12.65 | |
| RABModel=Whisper (LoRA)2026.06 | 21.27 | 12.87 | |
| ECAPA-TDNNModel=Whisper (LoRA)2026.06 | 21.27 | 10.54 | |
| Incremental SAT systemPercentile of Data Used=80%, Pseudo Labeling - Data Selection=Conf. Score, Pseudo Labeling - Incremental Iteration=4, SAT=✓, TTA=✓, Finetune Strategy=Semi-Sup.2026.06 | 21.5 | 13.32 | |
| i-vectorModel=Whisper (LoRA)2026.06 | 21.75 | 10.99 | |
| Semi-Supervised + SAT + TTAPercentile of Data Used=80%, Pseudo Labeling - Data Selection=Conf. Score, Pseudo Labeling - Incremental Iteration=x, SAT=✓, TTA=✓, Finetune Strategy=Semi-Sup.2026.06 | 22.13 | 12.87 | |
| Incremental SI systemPercentile of Data Used=80%, Pseudo Labeling - Data Selection=Conf. Score, Pseudo Labeling - Incremental Iteration=4, SAT=x, TTA=x, Finetune Strategy=Semi-Sup.2026.06 | 22.23 | 14.87 | |
| Semi-Supervised (Conf. Score)Percentile of Data Used=80%, Pseudo Labeling - Data Selection=Conf. Score, Pseudo Labeling - Incremental Iteration=-, SAT=x, TTA=x, Finetune Strategy=Semi-Sup.2026.06 | 22.4 | 13.54 | |
| Semi-Supervised (Random, Incr.)Percentile of Data Used=80%, Pseudo Labeling - Data Selection=Random, Pseudo Labeling - Incremental Iteration=4, SAT=x, TTA=x, Finetune Strategy=Semi-Sup.2026.06 | 22.49 | 14.54 | |
| Supervised (10%) + SAT + TTAPseudo Labeling - Incremental Iteration=x, SAT=✓, TTA=✓, Finetune Strategy=10% Sup.2026.06 | 22.86 | 13.32 | |
| x-vectorModel=Whisper (LoRA)2026.06 | 23.37 | 12.87 | |
| Semi-Supervised Baseline (Sys.3)Percentile of Data Used=100%, Pseudo Labeling - Data Selection=-, Pseudo Labeling - Incremental Iteration=-, SAT=x, TTA=x, Finetune Strategy=Semi-Sup.2026.06 | 23.62 | 13.32 | |
| Supervised (10%)Pseudo Labeling - Incremental Iteration=x, SAT=x, TTA=x, Finetune Strategy=10% Sup.2026.06 | 23.7 | 13.43 | |
| Conformer Transducer with Only Speech ContextsModel=Conformer Transducer2026.06 | 25.49 | 15.45 |