# duplex server / offline driver (same env as the Dynamic Accent Flash TTS: torch, CosyVoice deps)
fastapi
uvicorn[standard]
websockets
httpx
silero-vad          # MIT
fasttext-wheel      # GlotLID (cis-lmu/glotlid, Apache-2.0) for language ID
huggingface_hub
soundfile
librosa
numpy
# ASR worker env (separate venv is fine): omnilingual-asr>=0.2.0 (Apache-2.0), huggingface_hub, fastapi, uvicorn  (Dynamic ASR weights: Modularcomputing/dynamic-asr)
# LLM server: vllm