Text to speech

What the TTS layer does

TTS turns the language model's text output into audio the caller hears. It is one of three AI services on every Finn, alongside STT and the LLM.

You do not pick a TTS engine directly. You pick a voice, and the voice determines the engine. Voice selection lives on the Identity tab of the Finn editor. See Voice for the voice library itself.

The default

Every Finn is created with a default TTS service already set. It is tied to whichever voice you selected. If you never open the AI service configuration section, the default is what runs.

The AI service configuration section is marked advanced in the dashboard for a reason: changing TTS independently of voice is not something most accounts need. The documented guidance is to change it only if you have a specific reason.

What actually changes the audio

Three fields, in order of how much they matter.

Changing Language filters the voice list. If your current voice does not support the new language, you have to reselect. Not every voice supports every language.

Latency

TTS is part of the response latency the caller experiences. Premium voices may add a few hundred milliseconds. The voice library shows latency tiers per voice, so check the tier before committing to a voice on a latency-sensitive use case.

Latency also varies by language quality tier:

Above roughly 1.5s the conversation stops feeling natural. Callers start talking over the agent, which then feeds barge-in into STT and compounds the problem. If your Finn feels sluggish, the voice you picked is the first thing to check, before you start editing prompts.

When to change it

Change the voice, not the TTS service, in almost every case. Reasons to change voice:

• The tone does not match the use case. Calm and warm for healthcare, clear and direct for logistics, energetic for sales.
• You switched Language and the current voice does not cover it.
• Latency is too high and you are on a premium voice.
• Callers report the voice is hard to understand. This is common on Tier 3 languages, where the voice is clearly synthetic but understandable.

Reasons to touch the TTS service field itself are narrow, and if you cannot state yours in one sentence, leave it alone.

What TTS will not fix

TTS controls how words are spoken, not which words. If the agent says the wrong thing, that is the LLM and your prompt, not the voice. See Prompting.

Three common misattributions:

Custom vocabulary deserves a note. It is a list of words the speech recognizer should expect, and it materially improves how brand names, product SKUs, and unusual proper nouns are handled. If a caller-facing name comes out wrong, add it there before you go voice shopping.

Multilingual behavior

If you enable auto-detect multilingual mode, the Finn switches language mid-call. The documented cost is about one second of added latency on the first switch, and it only works reliably between Tier 1 languages. That latency is the switch itself, not steady-state TTS. See Multilingual.

For code-switching callers who mix two languages in one sentence, Tier 1 voices follow the caller's lead. Tier 2 and Tier 3 voices are not documented as handling this.

Dashboard, not API

Voice, Language, and TTS service are set in the Finn editor in the dashboard. There is no documented endpoint for changing the TTS service on a Finn. The voice library is readable via Voices API.

Changes apply to new calls. Calls already in flight finish on the old configuration, so a voice change during a live campaign produces a mixed set of recordings. Test the new voice on yourself first with Test on Call, which routes a real call to your phone. See Testing.

Next

• Voice — the voice library and how to pick.
• STT — the input side of the audio path.
• Multilingual — language switching and its costs.