Voice Comparison Studio
Generate and compare three voice models using the configured server voice.
CosyVoice 3 | Loaded; verify generation with a listening test |
whisper | Loaded on CPU |
whisper | Loaded on CPU |
ecapa | Loaded on CPU |
scoreq | Loaded on CPU |
dnsmos | Loaded on CPU |
Output language
Outputs keep their natural duration. ZeroGPU has queues and daily quotas. Evaluation measures output; it does not improve it.
TTFT is N/A because the adapters do not expose first-token events. FFE requires matching reference and generated sentences. Quality predictors are proxies, not human listening scores. Missing scores are never filled with zero.
Raters must score the same recordings for agreement. Ratings use local SQLite and are lost on container replacement unless suitable persistent storage is configured. Generated files expire after roughly two hours when later requests trigger cleanup.