Speech Recognition and Diarization on AMI IHM
19.21WERVibeVoice Single-pass baseline
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| VibeVoice Single-pass baselineChunk size=300s, Batch size=16, Total (s)=370.73, Speed-up=baseline2026.05 | 19.21 | 9.42 | 24.84 | 25.68 | |
| MURMUR: Chunked, no evictionChunk size=300s, Batch size=16, KV cache eviction strategy=None, Total (s)=100.8, Speed-up=3.68×2026.05 | 19.8 | 8.45 | 22.36 | 24.92 | |
| MURMUR: Output token evictionChunk size=300s, Batch size=16, KV cache eviction strategy=s4 w1024, Total (s)=97.6, Speed-up=3.79×2026.05 | 19.89 | 8.52 | 22.36 | 25.01 | |
| MURMUR: Both evictionChunk size=300s, Batch size=16, KV cache eviction strategy=Output and Speech eviction, Total (s)=96.1, Speed-up=3.85×2026.05 | 20.23 | 8.63 | 22.96 | 25.73 | |
| MURMUR: Speech token evictionChunk size=300s, Batch size=16, KV cache eviction strategy=d512 w1024, Total (s)=88.43, Speed-up=4.20×2026.05 | 20.41 | 8.77 | 23.19 | 25.29 | |
| WhisperXChunk size=300s, Batch size=16, Total (s)=38.81, Speed-up=9.55×2026.05 | 22.51 | 23.62 | 33.48 | 35.54 |