Voice and speech

Choosing a voice

Voice is set per Finn in the dashboard, not through the API at create time. Open the Finn, go to the Identity tab, pick the Language, then pick a voice from the filtered list. The voice library only shows voices that support the language you selected, so change the language first or the voice you want may not appear.

Voice coverage varies by language. English has the broadest selection (50+ voices across accents and genders). Spanish, French, and German have roughly 15-25 voices each with regional variants. Indian languages have 3-8 voices per language. Tier 3 languages typically have 1-3 voices. See supported languages for the tier list.

The templates in the gallery pair a voice with a use case. Those pairings are a reasonable starting point:

Pacing

Pacing is not a numeric slider. You control it through the Style & Guardrails field on the Identity tab, in plain language. The model follows those instructions for the whole call.

Instructions that work:

Speak in short, natural sentences. Mirror the caller's energy — calm if they're calm, faster if they're in a hurry. If the caller is upset, slow down and acknowledge them.

Length limits belong here too, because pacing and call length interact. A Finn told to keep a call under 90 seconds will compress its turns whether or not you asked it to speak faster:

Keep the call under 90 seconds unless they want to reschedule.

Two things go wrong with pacing instructions. First, if you write a long list of style rules, later rules get less weight than earlier ones — put the pacing rule near the top of the field. Second, a voice with an inherently slow delivery (Riley) will not become fast because you asked it to. Pick the voice for the baseline pace, then use Style & Guardrails for adjustment within that range.

Latency

Response latency depends on the language tier, not the voice:

Above roughly 1.5s the call stops feeling like a conversation. Callers start talking over the Finn because they assume the line dropped. If you are on a Tier 3 language and latency is your problem, there is no voice setting that fixes it. Consider a Tier 1 language if your callers are comfortable with one.

Multilingual auto-detect adds about one second of latency at the moment of the first language switch. That penalty lands once, on the first caller turn.

Interruption handling

Interruption behaviour is not exposed as a separate setting today. What you can do is instruct the Finn on how to behave when it is cut off, using Style & Guardrails, and keep its turns short enough that being cut off matters less. Long monologues are the main cause of callers talking over a Finn.

If a caller consistently talks over the Finn, check the transcript in call logs before changing the voice. The usual causes are a turn that ran too long, or a latency gap the caller read as silence.

Pronunciation

Use Custom Vocabulary on the Identity tab to fix words the Finn mispronounces or mishears. It takes a list of strings and works in any language, including mixed-language lists:

["Ola", "Zomato", "Razorpay", "Aadhaar", "PAN card", "GST"]

Add:

• Brand names, especially non-English ones with unusual spellings
• Product SKUs and model numbers
• Industry terminology
• Caller name spellings your callers commonly use

Custom vocabulary affects both directions: the Finn is more likely to say the word correctly and more likely to transcribe it correctly when a caller says it. Without it, an uncommon brand name will often come back in the transcript as a phonetically similar English word, which then breaks any post-call field that matches on that word.

Accents and code-switching

Tier 1 English handles most accents. Heavy regional accents raise the speech-to-text error rate, and there is no setting that compensates. Test with callers who sound like your real callers before launching.

Callers in some regions mix languages within one sentence. Tier 1 voices handle this and will follow the caller's lead. To make it explicit, add to Style & Guardrails:

Match the caller's speaking style — if they mix English and Hindi, you can mix too. Don't correct or comment on their language choice.

Custom vocabulary matters more than usual for code-switching callers, since a mixed-language sentence gives the transcriber less context per word.

What is not configurable

Language and voice cost the same per minute regardless of which you pick.

Next

• Prompting — how Identity, Style, and Guardrails fields interact.
• Multilingual — language list, tiers, and switching modes.
• Knowledge base — retrieval is language-agnostic, with caveats.