Voice service settings

Status: Current | Last reviewed: 2026-09-29

Open a voice or transcription node and select Properties. Integration profiles are limited to active, compatible profiles available in your tenant. The default profile is selected automatically; if there is only one compatible profile, it is selected. You can choose a different profile. When several profiles exist without a default, choose one explicitly.

Language lists readable names such as English (US) or French. The app saves the code required by the selected provider. Audio sample rates are managed by the channel and provider adapters, so there is no sample-rate setting to fill in.

For ElevenLabs → Scribe Realtime, select Scribe v2 Realtime from Model. Under Advanced → Key Terms, type a word or phrase and press Enter. You can paste several terms separated by commas. Each term appears separately; use its remove button to delete it. Names, products and specialist terms can help guide transcription. No JSON or brackets are required.

Properties → Auxiliary Services uses the same selectors for the private prompt voice, guided-flow transcription and small models. Defaults follow the agent's stack, tenant and provider configuration. Only your changes are saved as overrides. Choose Use defaults to inherit again. You can change the private voice even when its initial voice matches the main agent.

See Amazon Polly for language, compatible voices, engines and Play sample. If a preview fails, a message explains what to check before retrying.

ElevenLabs v4 and separate voices

To start with a separate ElevenLabs voice, choose Voice agent → Pick a stack → OpenAI Realtime + ElevenLabs v4. Select both integrations, then choose the voice on the ElevenLabs node. Realtime listens and prepares the reply; ElevenLabs speaks it. Keep your agent's procedures, tools and rules on the Realtime node.

For an existing text-generating pipeline, choose ElevenLabs → Text to Dialogue. Its model picker includes v4 Turbo (realtime) and v4 (quality). Refresh models checks your integration's current catalog. If discovery fails, the documented choices remain and an error explains that availability was not verified. A missing saved model stays visible until you choose a replacement. Play sample previews the selected model. V4 exposes Stability and Similarity; its older Style, Speed and Speaker Boost controls do not apply.

On the ElevenLabs node, Speech Formatting offers:

Use Advanced → Additional Delivery Guidance for preferences such as “Use a calm, professional voice and pause briefly before the reference number.” Formatting adds guidance to the generating model without another AI call. Listen to actual conversations to check the results; tags and prompting do not guarantee delivery or enforce business rules. Configure response guards separately when speech needs checking before playback.

You can also use speech recognition followed by a streaming text model and ElevenLabs, including Gemini text models. Current Google ADK Live models require native audio output. GPT-Live also retains native speech and is not offered as a text-only source for ElevenLabs. Changing models does not upgrade existing agents automatically.