Speech quality
ASR transcript fidelity is a comparative signal, not naturalness, expressiveness or speaker identity. Listening review remains necessary.
| Run | Engine | Checkpoint | Delivery | Judged | Exact % | Mean similarity % | Minimum % | Repeat consistency % | Errors | ASR model |
|---|---|---|---|---|---|---|---|---|---|---|
| 20261025-050404-08222176 | faster | 0.6B-CustomVoice | complete | 100 | 83.00 | 99.36 | 90.91 | 55.00 | 0 | Qwen/Qwen3-ASR-0.6B-hf |
| 20261025-050404-08222176 | faster | 0.6B-Base | complete | 30 | 20.00 | 97.94 | 87.89 | 0.00 | 0 | Qwen/Qwen3-ASR-0.6B-hf |
| 20261025-053632-9d5b8428 | faster | 1.7B-CustomVoice | complete | 100 | 81.00 | 99.00 | 72.94 | 55.00 | 0 | Qwen/Qwen3-ASR-0.6B-hf |
| 20261025-053632-9d5b8428 | faster | 1.7B-Base | complete | 30 | 6.67 | 96.75 | 83.92 | 10.00 | 0 | Qwen/Qwen3-ASR-0.6B-hf |
| 20261025-053632-9d5b8428 | faster | 1.7B-VoiceDesign | complete | 100 | 90.00 | 99.58 | 90.91 | 60.00 | 0 | Qwen/Qwen3-ASR-0.6B-hf |
See DETAILS.md for the latest run and methodology.