How gpustack/vox-box works
File structure
• Top-level code lives under vox_box/:
• main.py – CLI entry point
• cmd/ – subcommands (start, version)
• server/ – HTTP server implementation: server.py (launcher), app.py (FastAPI/Flask app), routers.py (route definitions), model.py (API data models)
• backends/ – abstract base classes and concrete adapters for each STT/TTS engine
• estimator/ – orchestration layer that chooses and invokes the correct backend for a request
• downloader/ – logic to fetch and cache models from Hugging Face or ModelScope
• config/ – configuration loader for CLI flags and environment settings
• utils/ & logging.py – cross-cutting utilities (file I/O, logging, compatibility)
• third_party/ – bundled external code for CosyVoice and Dia
• No relational or document database; persistent state is the on-disk model cache
• No separate frontend; clients are CLI and any HTTP client
• CI/build scripts in hack/ and GitHub Actions workflows, but these are orthogonal to runtime architecture
README for additional insights
• Confirms REST endpoints: /v1/audio/speech, /v1/audio/transcriptions, model listing, health check
• Describes dependencies (Python ≥3.10, NVIDIA CUDA/cuDNN)
• Mentions caching models under --data-dir