Vocence Miner03

Setup

• fish_speech/models/text2semantic/inference.py and fish_speech/models/dac/inference.py from a clone of Fish Speech.

Config (vocence_config.yaml)

• runtime.fish_speech_root (or FISH_SPEECH_ROOT) — path to the Fish Speech repo
• runtime.fish_checkpoint_dir — S2 Pro weights (default bakes to /app/checkpoints/s2-pro in the Chute image)
• runtime.fish_half / fish_compile — optional performance flags
• generation.max_seconds — cap output length

Notes

• S2 Pro is VRAM-heavy; production serving generally expects high-memory GPUs.
• Chutes TEE: Fish scripts run in-process by default (runpy). Spawning a separate python child fails with aegis[integ] FATAL: security violation (secure authorization fd not valid). For local debugging with isolated subprocesses, set VOCENCE_FISH_USE_SUBPROCESS=1.
• Keep the Vocence wrapper unchanged except approved template variables.