Models and stacks

Model ID format

Every model is identified by a provider/model string:

deepgram/nova-3
openai/gpt-4.1-mini
cartesia/sonic-3

This is the exact string cost tracking prices. When attach()/guard() wraps a plugin, VoiceGateway reads its provider and model attributes off the live instance (f"{provider}/{model}") and passes that straight to voice-prices. Nothing is parsed or stripped from it.

Language and voice selection belong to your framework’s native constructor kwargs (language="en" on a Deepgram STT plugin, voice_id="..." on a Cartesia TTS plugin), not to the model string. If a colon-suffixed value like nova-3:en ends up as the plugin’s model, that suffix becomes part of the literal model_id VoiceGateway records, and the pricing lookup for that exact string fails for most cloud providers (cost records as unpriced). Local/self-hosted models (local/*, ollama/*) are unaffected: they price at $0 regardless of what follows the model name, which is why an Ollama tag (ollama/llama3.2:3b) or a Piper voice (local/piper:en_US-lessac-medium) is safe to write as-is.

Registering custom model aliases

Register aliases under models in voicegw.yaml. Aliases surface in the dashboard and CLI as friendlier names; they group models by modality with an optional default_voice for TTS entries.

models:
  stt:
    fast-stt:
      provider: deepgram
      model: nova-3
  llm:
    reasoning:
      provider: anthropic
      model: claude-sonnet-4-5
  tts:
    narrator:
      provider: cartesia
      model: sonic-3
      default_voice: narrator-male

Field	Required	Description
`provider`	yes	Provider identifier (e.g. `deepgram`, `anthropic`)
`model`	yes	Model name at the provider
`default_voice`	no	Display-only default voice for TTS aliases

Models can also be registered through the dashboard or, if voicegw mcp is running, through MCP tool calls from your IDE. Dashboard/MCP-registered models persist in SQLite and merge with YAML at startup (a YAML entry with the same id wins).

Model reference

STT

Model ID	Provider	Notes
`deepgram/nova-3`	Deepgram	Best cloud STT accuracy
`deepgram/nova-2`	Deepgram	Lower cost alternative
`openai/whisper-1`	OpenAI	OpenAI-hosted Whisper
`groq/whisper-large-v3`	Groq	Fast Whisper via Groq
`assemblyai/universal-2`	AssemblyAI	High accuracy, single tier
`local/whisper-large-v3`	Whisper (local)	Best local STT; always $0

LLM

Model ID	Provider	Notes
`openai/gpt-4.1-mini`	OpenAI	Good cost/quality balance
`openai/gpt-4.1`	OpenAI	Best quality
`anthropic/claude-sonnet-4-5`	Anthropic	Strong reasoning
`anthropic/claude-haiku-4-5`	Anthropic	Fast and cheap
`groq/llama-3.3-70b-versatile`	Groq	Fast open-source LLM
`ollama/llama3.2:3b`	Ollama (local)	Always $0; tag after `:` passes through untouched

TTS

Model ID	Provider	Notes
`cartesia/sonic-3`	Cartesia	Low-latency streaming
`openai/tts-1`	OpenAI	Fast cloud TTS
`openai/tts-1-hd`	OpenAI	High quality cloud TTS
`elevenlabs/eleven_multilingual_v2`	ElevenLabs	29 languages
`deepgram/aura-asteria-en`	Deepgram	Deepgram TTS
`local/kokoro`	Kokoro (local)	Always $0
`local/piper:en_US-lessac-medium`	Piper (local)	Always $0; voice id after `:`

Stacks

A stack is a named bundle mapping one name to an STT, LLM, and TTS model id: exactly those three keys, nothing else.

stacks:
  premium:
    stt: deepgram/nova-3
    llm: anthropic/claude-sonnet-4-5
    tts: cartesia/sonic-3
  budget:
    stt: groq/whisper-large-v3
    llm: groq/llama-3.3-70b-versatile
    tts: local/piper:en_US-lessac-medium

Reference a stack from a project with default_stack: premium (see Projects).

A stack is a dashboard and documentation hint only. Nothing reads default_stack at runtime to construct a provider. Pick the model ids from your chosen stack and pass them to your framework’s native provider constructors yourself, then wrap with attach()/guard():

from livekit.agents import AgentSession
from livekit.plugins import deepgram, anthropic, cartesia
from voicegateway import attach

session = AgentSession(
    stt=deepgram.STT(model="nova-3"),
    llm=anthropic.LLM(model="claude-sonnet-4-5"),
    tts=cartesia.TTS(model="sonic-3"),
)
attach(session, project="acme")

See Projects for default_stack and budget configuration. See attach() for wiring native providers so cost tracking sees them. See voicegw.yaml reference for the full config file shape.

Instrument

Cost tracking

Conversations

Examples

Model ID format

Registering custom model aliases

Model reference

STT

LLM

TTS

Stacks

​Model ID format

​Registering custom model aliases

​Model reference

​STT

​LLM

​TTS

​Stacks

Model ID format

Registering custom model aliases

Model reference

STT

LLM

TTS

Stacks