Trost

trost is a clanker — the fleet's OpenAI-compatible API for VibeVoice TTS with an integrated voice-mode pipeline (ASR → LLM → TTS).

Stack: Python. Private repo. Targeted at RTX 3090.

Supports two model variants:

• VibeVoice 7B — full quality, ~19 GB VRAM (~7 GB with bnb-4bit).
• VibeVoice-Realtime 0.5B — low latency (<500 ms), ~2.5 GB VRAM.

Features

• OpenAI-compatible TTS — /v1/audio/speech.
• Voice-mode pipeline — ASR (Whisper) → LLM → TTS.
• Live voice chat with microphone.
• Voice Activity Detection — Silero VAD for automatic turn-taking.
• Multi-speaker dialogue synthesis.
• Zero-shot voice cloning.
• WebSocket streaming for real-time voice with VAD.
• GPU acceleration.

Endpoints

OpenAI-compatible:

• POST /v1/audio/speech
• GET /v1/audio/voices
• POST /v1/audio/transcriptions (Whisper)
• GET /v1/audio/models

VibeVoice-specific:

• POST /v1/vibevoice/generate — advanced TTS with full options.
• POST /v1/vibevoice/clone — voice cloning with file upload.
• POST /v1/vibevoice/multi-speaker — dialogue synthesis.
• GET /v1/vibevoice/health.

Voice pipeline:

• POST /v1/voice/pipeline — full voice mode (ASR → LLM → TTS).

Batteries-included via docker compose up -d (Whisper + LLM + TTS).

Sources

• README

See also

• discord-bot — parallel voice stack for Discord.
• ears — sibling ASR daemon.
• x-vector — voice-embedding workbench that feeds TTS speaker embeddings.