Model server

Run STT, TTS, and LLM on your own GPUs behind one gateway.

An optional, self-contained subsystem for running speech and language models on your own hardware. One gateway on :8100 fronts three slots — STT, TTS, LLM — and each slot runs whichever model you selected, without the gateway knowing which one.

The model server is optional. VoicEra runs fine against cloud providers alone. Deploy this when data residency, per-minute cost, or Indic-language coverage makes self-hosting worth the GPUs.

Start here

Page	What it answers
Overview	What it is, what it replaces, and its current state.
Slots and models	The slot/model split, and switching models with one line.
Gateway API	Every endpoint, including WS /v1/asr/ws.

The models

Page	Covers
STT models	indic-conformer, indic-transcribe, and partial transcripts versus streaming.
TTS models	indic-parler, orpheus, indic-mio, and audio format negotiation.
LLM models	qwen3.5-4b and the model-id agreement across files.

Operating it

Page	Covers
Adding a model	The container contract, and the two steps to add one.
Running on GPUs	Device selection, MPS, shared HuggingFace cache, disk.

Honest status

Two gaps are documented rather than hidden. The LLM slot has not yet been run on real hardware, and a full call routed from the voice runtime through a self-hosted model has not been verified end to end. Both are called out on the pages concerned.

Related

• Self-hosted models — wiring the model server to the runtime
• Provider registry — how agents select a provider