Skip to main content

Agents

Voice and speech

Choosing a voice, matching it to the language, and testing it.

Choosing a voice

Each Finn speaks with one voice. Voices come from text-to-speech providers; there is no fixed Finn voice roster.

In the dashboard, click Edit on the Finn's builder page, then Call flow → Identity → Voice persona. The Voice persona dialog lists voices from these providers: ElevenLabs, Cartesia, Azure, Sarvam and Inworld.

Narrow the list with the Provider, Gender, Accent, Use case and Language filters, or search by name. Play a voice's sample before you pick it; the sample button next to Voice persona plays it again later.

Picking a voice also sets the Finn's text-to-speech service to that voice's provider, so you don't choose the provider separately. See agents-tts.

Setting the voice through the API

Through the API, voice is required when you create a Finn. Look up a voice_id with api-voices, which lists every voice your organization can use, the same as the dashboard's voice picker, then pass it as voice on api-finns. The API checks the voice and puts the Finn on that voice's provider; a voice that isn't one of yours returns 400 invalid_request.

Match the voice to the language

The Finn speaks its Default language in the voice you choose, so the voice must handle that language well. Use the Language filter to find voices for your Finn's language, then confirm on a test call. See agents-multilingual.

Pacing and style

How the Finn talks, such as sentence length and how it responds to a hurried or upset caller, comes from your instructions, not the voice. Put it in the system prompt:

Speak in short, natural sentences. If the caller is in a hurry, get to the point. If the caller is upset, slow down and acknowledge them before answering.

Short turns matter most. Long answers are the main reason callers talk over a Finn. Pick the voice for its baseline sound, then shape delivery with instructions. See agents-prompting.

The Agent handbook tab also has delivery toggles, such as Natural filler words and Speech normalization, which reads numbers and money naturally. See agents-guardrails.

Testing a voice

A sample tells you how a voice sounds, not how it sounds on a phone line reading your script. Before you deploy, place a test call with Test → Phone call and listen for:

  • how the welcome message sounds in that voice;
  • how it reads numbers, dates and your brand name;
  • whether the pace suits your callers.

If a caller keeps talking over the Finn, check the transcript in call-logs before you change the voice. The usual cause is a turn that ran too long.

Next