Skip to main content

Agents

Multilingual calls

Detecting and switching language mid-call.

When to use multilingual mode

A Finn speaks one language for the whole call unless you configure otherwise. Multilingual mode lets it detect the caller's language from their first response and switch. Use it when one phone number serves callers who speak different languages and you do not know which in advance.

If you already know the caller's language (separate DIDs per language, or a CRM property on the contact), do not use detection. Run separate Finns instead. Detection costs latency and can misfire; a known language costs nothing.

Option A — auto-detect at call start

The Finn opens in its primary language, listens to the caller's first response, and switches if the detected language differs.

To enable, in the dashboard:

  1. Open the Finn, go to the Identity tab.
  2. Set Language to the primary language. This is what the welcome message is spoken in.
  3. Set Multilingual mode to Auto-detect.
  4. Select the candidate languages. You can select up to 3.
  5. Pick a voice that supports every candidate language. See voice.
  6. Save.

There is no API field for multilingual mode. It is dashboard-only.

What goes wrong:

ProblemCauseWhat to do
Noticeable pause after the caller's first sentenceThe switch adds about 1 second of latencyExpected. Keep the welcome message short so the pause lands early
Wrong language pickedCaller answered with a one-word reply, or a Tier 2/3 language was a candidateDetection is reliable only between Tier 1 languages. Restrict candidates to Tier 1
Finn switches but the voice sounds wrongThe selected voice does not cover the second languagePick a voice from the filtered list that supports all candidates
Caller code-switches and the Finn keeps flippingDetection reacts to mixed-language inputUse code-switching guidance below instead of multilingual mode

Tier 1 languages are English, Spanish, French, German, Portuguese, Italian, Japanese, Mandarin Chinese, Hindi, Arabic, and Korean. Everything else falls into Tier 2 or Tier 3 and should not be a detection candidate.

Option B — explicit language menu

The Finn opens with a bilingual prompt and the caller states a choice:

For English, say English. Para español, diga español.

Build this as a Workflow: a Decision node keyed on the language detected at the first Ask node. This is explicit and it does not misfire. It reads like an IVR, which is the tradeoff. Choose it where multilingual disclosure is a legal requirement in your jurisdiction. See compliance.

Option C — separate Finns per language

Run one Finn per language and route before the call reaches an agent. Route inbound on the dialed number (one DID per language) via inbound routing, or on a contact property from your CRM for outbound.

Each Finn gets its own prompt, voice, knowledge base, and post-call fields, tuned for that language. The cost is duplicate maintenance: every prompt change has to be made twice. This is the right choice when each language audience is large enough to own separately.

Code-switching

Code-switching is a caller mixing two languages inside one sentence — Hinglish, Spanglish, French-Arabic. This is not the same as switching languages, and multilingual mode is the wrong tool for it. Tier 1 voices follow the caller's lead already.

To make it explicit, add to Identity → 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.

Add the non-primary-language terms callers will use to Identity → Custom Vocabulary. Brand names and local terms are what STT gets wrong first:

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

Custom vocabulary applies across languages, so one list covers both sides of a code-switched call. See prompting.

Latency by tier

Round-trip response latency varies by language tier. Above roughly 1.5 seconds a call stops feeling like a conversation.

TierLatency
Tier 1Under 800ms
Tier 2800-1000ms
Tier 31000-1300ms

Auto-detect adds about 1 second on top of this, once, at the first switch. A Tier 3 language plus a detection switch will be audible to the caller.

Knowledge base and language

The knowledge base is language-agnostic. You can upload documents in any language and the Finn retrieves relevant content regardless of the source language, translating as needed in its response.

Cross-language retrieval works but is less reliable than same-language retrieval. For a multilingual Finn, upload knowledge in the primary language and accept that answers in the second language are translated at call time. If accuracy in both languages matters, Option C is the safer structure.

What multilingual mode is not

It is not live translation. The Finn speaks one language at a time and does not relay between two parties who speak different languages. Auto-detect is the closest available behavior.

Pricing does not vary by language. All languages cost the same per minute.