This repository serves as the official model host for Vox, a real-time, local-first voice-to-voice system. It contains specialized models for Voice Activity Detection (VAD), Speech-to-Text (STT), Large Language Models (LLM), and Text-to-Speech (TTS).
Directory Structure
The structure of this repository exactly mirrors the runtime expectations of the Vox backend:
.
├── manifest.json # Single source of truth for all models
├── llm/
│ └── gemma4/ # Large Language Models (GGUF)
├── stt/
│ └── qwen3-asr/ # Speech-to-Text (ONNX)
│ └── tokenizer/ # STT Tokenizer configs
├── tts/
│ ├── kokoro/ # English TTS (Kokoro ONNX)
│ └── piper_hi/ # Hindi TTS (Piper ONNX)
└── vad/
└── ten_vad.onnx # Voice Activity Detection (ONNX)
Manifest
The manifest.json file in the root directory provides metadata for automated management, including:
Relative file paths
Exact byte sizes
SHA256 hashes for integrity verification
Archive markers for compressed assets (e.g., espeak-ng-data)
Usage
These models are intended to be downloaded and managed by the Vox application runtime. For manual use, ensure you have Git LFS installed to correctly retrieve the large model weights.
