feat: add GET /debug/gpu to report live VRAM/RAM usage

perf: don't load Whisper ASR by default (faster cold starts)

fix: align /tts/clone class_temperature default to 0.3

Update app.py

Revert "feat: optional server-side background ambience mix"

feat: optional server-side background ambience mix

feat: expose class_temperature/position_temperature in TTSRequest

feat: add GET /debug/voices-dir to diagnose persistent storage

refactor: use torch.save/load instead of pickle for voice prompts

feat: persist voice registry to /data/voices across restarts

Add clone-once voice registry (POST /voices + voice_id reuse)

Match upstream pins

Pin torch==2.8.0 + transformers==5.3 to fix higgs_audio_v2_tokenizer import

Self-hosted OmniVoice TTS on T4 (FastAPI, no proxy)

initial commit