PowerBeef02/Qwen3-TTS-12Hz-1.7B-CustomVoice-8bit

Restore the fp32 speech tokenizer (sha256 836b7b35…, 682,293,092 bytes)

Add required AI-marking generator (EU AI Act Article 50): AudioSeal wm16 fused fp16 export (MIT, facebookresearch/audioseal @ cc2700db)

codec f16 (artifactVersion 2026.08.01.1): speech tokenizer F32->F16 (492 tensors cast, final two decoder layers kept F32 per T-Mimi), -341 MB. Validation: Vocello benchmarks/OPTIMIZATION.md SR.

Vocello 2.2 artifact: 8-bit text embedding (from mlx-community@41d3337e8b7f)

initial commit