Languages & Detection
Automatic language detection, expected-language hints, and multilingual conversations.
Kalima recognizes speech in 60+ languages and figures out what's being spoken for you, automatically. You don't have to pick a language before you start, and even conversations that switch back and forth between languages are transcribed correctly. This page explains how detection works, how to give Kalima a helpful nudge with expected-language hints, and how multilingual sessions appear in your transcript.
Automatic language detection
Whenever you record live or upload a file, Kalima listens to the audio and identifies the language being spoken, no setting required. The transcript simply appears in the right language.
This works the same way everywhere transcription happens:
- Live transcription, words appear in the detected language as you speak.
- Uploaded files, the recording's language is detected during processing and reflected in the finished transcript.
Detection is guided, not forced. Kalima makes its best judgement from the audio itself, and you can steer it toward the right answer with expected-language hints when you already know what will be spoken.
You never have to choose a language to get a transcript. Auto-detection is always on. Hints are an optional accuracy boost, not a requirement.
Multilingual and code-switching sessions
Plenty of real conversations slide between languages, a bilingual meeting, a language class, an international call where people drop into their native tongue and back. Kalima handles this in a single transcript.
When the spoken language changes mid-conversation, Kalima detects the switch and labels each part of the transcript with a small language badge showing which language that segment is in. A single session can contain several languages, each marked where it occurs, so a mixed-language record stays clear and readable.
Where language badges appear
Per-segment language badges show up directly in the transcript display in the Studio, next to the relevant lines. They make it easy to scan a multilingual transcript and see at a glance who spoke which language and where the conversation switched.
For bilingual sessions, pair detection with Live Translation so everyone can follow along. One-way translation renders everything into a single language, while two-way translates between two languages in both directions.
Expected languages: give Kalima a hint
If you already know which languages will come up, you can tell Kalima in advance by adding expected languages (also called language hints). This guides recognition toward those languages, which noticeably improves accuracy, without locking other languages out.
Hints are especially worth adding when:
- The audio is heavily accented and you want recognition tuned to the right language.
- The session is reliably bilingual (for example, you know it will be English and Spanish).
- You're transcribing a less common language that might otherwise be mistaken for a similar-sounding one.
Adding hints to a live session
Adding hints to a file upload
Best practice: add a hint only for languages that are actually in the recording. Listing languages that aren't present can pull detection in the wrong direction, fewer, accurate hints beat a long, speculative list.
How many hints can I add?
You can add more than one expected language to a session. The number allowed per session is governed by your plan's languages-per-session limit, which is currently unlimited on standard plans. The interface shows a running count as you add languages and displays a notice if a cap is ever reached.
A sample of supported languages
Kalima's picker offers 60+ languages spanning the major world languages and many regional ones. This is a representative selection, not the full list, to give you a feel for the coverage:
| Region | Examples |
|---|---|
| Western Europe | English, German, French, Spanish, Italian, Portuguese, Dutch |
| Northern & Eastern Europe | Polish, Swedish, Norwegian, Danish, Finnish, Czech, Ukrainian, Russian |
| Middle East & Africa | Arabic, Hebrew, Turkish, Persian, Swahili |
| South & Southeast Asia | Hindi, Bengali, Tamil, Urdu, Thai, Vietnamese, Indonesian, Malay |
| East Asia | Mandarin Chinese, Cantonese, Japanese, Korean |
You'll find the complete, searchable list in the Languages step when you set up a session or upload a file.
Tips for the most accurate detection
A few habits make a real difference to how reliably Kalima identifies and transcribes your languages:
- Add expected-language hints whenever you know what will be spoken.
- Use clean audio, a good microphone and the right audio quality mode help the engine hear distinct sounds clearly, which improves both detection and accuracy.
- Add session context for names, jargon, and proper nouns so specialized terms land correctly alongside the right language.
- Avoid very short clips when you can, a longer sample gives detection more to work with, especially for similar-sounding languages.
Detection confidence can dip on very short clips or noisy audio. If a language looks wrong, re-running with an expected-language hint (and cleaner audio) usually fixes it.
Frequently asked questions
No. Language is detected automatically, and the transcript handles mid-conversation language switches. For better accuracy, you can optionally add expected-language hints if you already know what will be spoken.
60+ languages are available in the picker, covering the major world languages and many regional ones. Any of them can be detected automatically or added as an expected-language hint.
Yes. Kalima detects language switches within a single session and labels each part of the transcript with a language badge, so bilingual meetings, language classes, and international calls are transcribed correctly in one transcript.
No. Hints guide recognition toward the languages you expect, but the engine still detects the languages actually spoken. That's why it's best to list only languages that are really present in the audio.
You can add more than one. The per-session count follows your plan's languages-per-session limit, which is currently unlimited on standard plans. The interface shows a counter as you add languages and a notice if any cap is reached.
This usually happens with very short clips, noisy audio, or two similar-sounding languages. Add an expected-language hint for the correct language, improve your audio quality, and re-record or re-upload. Adding session context for names and terms can also help.