Datasets:
Gemini rank-32 CFG and checkpoint listening audio
The original High-8654 / early update-13,600 comparison remains under audio/, with its original RESULTS.json and PROTOCOL.json unchanged. Its complete page is preserved in the linked Space as early-13600.html.
The new checkpoint study is isolated under sweep-20261104/: generated evaluation MP3s, per-outcome raw scoring metadata, exact matched and showcase recipes, source hashes, approved fixed female reference, and a completed selection record when available. Audio and failures appear incrementally. See that prefix's protocol and results for completion counts and metric definitions. Wordless examples have null WER; failed generations have metadata and no broken audio placeholder.
Selection is a development-set diagnostic combining frozen emotion/style target CDF, content enjoyment and genuineness. WER, frame caps and nonfinite values are reported independently. Additional five-language demonstrations use the same fixed reference and CFG 2.5; they do not select the checkpoint. No model weights or training corpus are published here.
Listening Space: https://huggingface.co/spaces/TTS-AGI/humaneness-voice-gemini-r32-cfg-comparison
- Downloads last month
- 715