Automatic Speech Recognition on LibriSpeech C&W attack
0WERCTC-FP32
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| CTC-FP32Inference precision=FP322026.03 | 0 | 0 | 18.54 | |
| CTC-FP16Inference precision=FP162026.03 | 0 | 0 | 17.94 | |
| CTC-BF16Inference precision=BF162026.03 | 0 | 0 | 17.53 | |
| seq2seq-FP32Inference precision=FP322026.03 | 0 | 0 | 13.34 | |
| seq2seq-BF16Inference precision=BF162026.03 | 0 | 0 | 17.23 | |
| Transformer-FP32Inference precision=FP322026.03 | 0 | 0 | 28.42 | |
| Transformer-FP16Inference precision=FP162026.03 | 0 | 0 | 28.42 | |
| Transformer-BF16Inference precision=BF162026.03 | 0 | 0 | 28.27 | |
| Whisper-FP32Inference precision=FP322026.03 | 0 | 0 | 26.37 | |
| Whisper-BF16Inference precision=BF162026.03 | 0 | 0 | 26.3 | |
| seq2seq-FP16Inference precision=FP162026.03 | 0.41 | 1 | 15.89 | |
| CTC-FP16Inference precision=FP322026.03 | 3.29 | 15 | 17.94 | |
| CTC-FP32Inference precision=FP162026.03 | 6.38 | 28 | 18.54 | |
| PVP (CTC-FP16)Inference precision=Random2026.03 | 10.29 | 34.5 | 17.94 | |
| Whisper-FP16Inference precision=FP162026.03 | 10.36 | 8 | 31.86 | |
| PVP (CTC-FP32)Inference precision=Random2026.03 | 10.93 | 38.9 | 18.54 | |
| PVP (CTC-BF16)Inference precision=Random2026.03 | 12.59 | 44.8 | 17.53 | |
| PVP (seq2seq-FP32)Inference precision=Random2026.03 | 18.5 | 41.2 | 13.34 | |
| CTC-BF16Inference precision=FP162026.03 | 18.93 | 67 | 17.53 | |
| PVP (Transformer-FP16)Inference precision=Random2026.03 | 19.34 | 32.8 | 28.42 | |
| CTC-BF16Inference precision=FP322026.03 | 19.55 | 70 | 17.53 | |
| Transformer-BF16Inference precision=FP162026.03 | 19.75 | 30 | 28.27 | |
| seq2seq-FP32Inference precision=FP162026.03 | 22.02 | 48 | 13.34 | |
| Transformer-FP16Inference precision=FP322026.03 | 22.63 | 40 | 28.42 | |
| PVP (Transformer-FP32)Inference precision=Random2026.03 | 23.85 | 31.3 | 28.42 | |
| Transformer-FP32Inference precision=FP162026.03 | 24.9 | 35 | 28.42 | |
| CTC-FP32Inference precision=BF162026.03 | 25.72 | 83 | 18.54 | |
| CTC-FP16Inference precision=BF162026.03 | 25.93 | 83 | 17.94 | |
| PVP (Whisper-FP16)Inference precision=Random2026.03 | 27.91 | 27.5 | 31.86 | |
| Whisper-FP16Inference precision=FP322026.03 | 28.69 | 31 | 31.86 | |
| PVP (seq2seq-BF16)Inference precision=Random2026.03 | 33.79 | 64.1 | 17.23 | |
| seq2seq-FP32Inference precision=BF162026.03 | 34.16 | 77 | 13.34 | |
| Transformer-FP16Inference precision=BF162026.03 | 35.19 | 56 | 28.42 | |
| PVP (Transformer-BF16)Inference precision=Random2026.03 | 37.45 | 55.2 | 28.27 | |
| PVP (seq2seq-FP16)Inference precision=Random2026.03 | 40.58 | 66.2 | 15.89 | |
| Whisper-FP16Inference precision=BF162026.03 | 44.42 | 43 | 31.86 | |
| Transformer-FP32Inference precision=BF162026.03 | 44.44 | 57 | 28.42 | |
| PVP (Whisper-FP32)Inference precision=Random2026.03 | 47.25 | 43.8 | 26.37 | |
| seq2seq-BF16Inference precision=FP322026.03 | 50.41 | 98 | 17.23 | |
| seq2seq-BF16Inference precision=FP162026.03 | 51.85 | 98 | 17.23 | |
| PVP (Whisper-BF16)Inference precision=Random2026.03 | 52.09 | 50.6 | 26.3 | |
| seq2seq-FP16Inference precision=FP322026.03 | 56.38 | 95 | 15.89 | |
| Transformer-BF16Inference precision=FP322026.03 | 58.44 | 86 | 28.27 | |
| seq2seq-FP16Inference precision=BF162026.03 | 62.35 | 97 | 15.89 | |
| Whisper-BF16Inference precision=FP322026.03 | 63.35 | 63 | 26.3 | |
| Whisper-FP32Inference precision=BF162026.03 | 66.14 | 61 | 26.37 | |
| Whisper-FP32Inference precision=FP162026.03 | 79.48 | 72 | 26.37 | |
| Whisper-BF16Inference precision=FP162026.03 | 92.63 | 88 | 26.3 |