Automatic Speech Recognition on In-house ASR
6.63AVG ScoreConcat-MLP
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Concat-MLPCompression Ratio (R)=8, Compression Method=Adapter-level2026.07 | 6.63 | — | 4.08 | 7.63 | 6.17 | 8.63 | |
| Window Q-FormerCompression Ratio (R)=8, Compression Method=Adapter-level2026.07 | 6.5 | — | 4.34 | 7.59 | 5.08 | 8.98 | |
| Concat-MLPCompression Ratio (R)=4, Compression Method=Adapter-level2026.07 | 6.29 | — | 4.66 | 7.45 | 5.14 | 7.91 | |
| Window Q-FormerCompression Ratio (R)=4, Compression Method=Adapter-level2026.07 | 6.23 | — | 4.91 | 7.39 | 4.98 | 7.65 | |
| HS CompressionCompression Ratio (R)=8, Compression Method=In-LLM, l0=52026.07 | 6.22 | — | 4.06 | 7.57 | 4.77 | 8.47 | |
| SpeechKVCompression Ratio (R)=8, Compression Method=In-LLM, l0=52026.07 | 5.95 | — | 3.95 | 7.27 | 4.45 | 8.14 | |
| HS CompressionCompression Ratio (R)=4, Compression Method=In-LLM, l0=52026.07 | 5.79 | — | 3.62 | 7.15 | 4.67 | 7.73 | |
| MLP (Baseline)2026.07 | 5.78 | — | 3.59 | 7.58 | 4.26 | 7.67 | |
| SpeechKVCompression Ratio (R)=4, Compression Method=In-LLM, l0=52026.07 | 5.71 | — | 3.83 | 7.09 | 4.52 | 7.42 | |
| Whisper-L-v32026.07 | 4.91 | — | 3.4 | 6.83 | 4.58 | 4.82 | |
| Phi4-mm2026.07 | 4.89 | — | 3.13 | 6.21 | 3.88 | 6.33 |