agenthub · output.voice
Voice Output
Speak text aloud. It takes an upstream answer and synthesises a clip the chat surface plays.
Why the node earns its place
The natural terminal for a voice agent: an LLM writes the answer, this says it. Only voice models appear in the dropdown — those whose deployment actually exposes speech synthesis — so you cannot select a model that will fail at call time.
Instructable models take direction. The instructions field is where tone, pace and emotion go, which is the difference between a correct answer and one that sounds like the brand.
How it works at run time
- 1Take the text from the wired port, or from a wired message object.
- 2Synthesise through the speech capability in the requested format, with the chosen voice and any instructions.
- 3Return an assistant message carrying a voice marker, the audio data URI and the spoken text; the surface renders a player.
What you wire
Inputs
- text *string
- messageobject
- instructionsstring
Outputs
Nothing handed onward.
Configuration
| Key | Type | Default & options | What it does |
|---|---|---|---|
| model * | string | choices from model-library:voice_models | Voice (text-to-speech) model to synthesize with |
| voice | string | — | Voice name for the selected model (provider-specific, e.g. OpenAI: alloy/nova, Gemini: Kore/Puck). Leave blank for the model's default. |
| format | string | default "mp3" mp3 · wav · opus | Audio format |
| instructions | string | — | How the voice should speak — tone, emotion, pace (gpt-4o-mini-tts and other instructable models). E.g. 'Warm, patient kindergarten teacher: cheerful, slow and clear, with little pauses.' Ignored by plain TTS models. |
What usually goes wrong
Watch for this
Voice names are provider-specific — an OpenAI voice name means nothing to Gemini. Leave it blank for the model's default rather than guessing. Plain, non-instructable models ignore instructions silently rather than erroring.
Behaviour & provenance