Most transcription software was built around a single unstated assumption: the speaker uses one language throughout the recording, and that language is probably English. This assumption works for a slice of the global professional world. It fails for everyone else — and the “everyone else” category is enormous.
Researchers studying dialect contact. Journalists whose sources switch to English for the technical points. Lawyers, consultants, and therapists with multilingual clients. Oral historians recording diaspora families. For all of them, a transcription tool that handles one language at a time — or that collapses when a speaker shifts languages — is not a professional tool, it is a toy.
Transcribes 98 languages on your laptop — all on-device. Accuracy varies by language — the list is on /transcribe. This guide explains how multilingual transcription works in Kajo and when it is the right tool.
What code-switching is and why it matters
Code-switching is the linguistic practice of alternating between two or more languages within a single conversation or even a single sentence. It is not a sign of linguistic confusion or incomplete language learning — it is a sophisticated, systematic communicative practice used by bilingual and multilingual speakers around the world.
Examples from natural speech:
- A Spanish-English bilingual consulting client: “Me siento muy overwhelmed lately, you know what I mean? Es como que everything is too much.”
- An Arabic-English business professional: “القرار كان clear من البداية — we need to pivot the strategy.”
- A French-English academic: “L’analyse shows that the hypothesis was correct, mais les données sont still ambiguous.”
In each case, the speaker moves naturally between languages. A transcription tool that processes only one language will either transcribe the non-dominant language as garbage text or skip those sections entirely. The result is a transcript with holes — precisely where the speaker’s most authentic expression may have occurred.
How Kajo handles multiple languages
One language decision per recording. Kajo’s engine makes one language decision per recording — automatically from the opening, or the language you set in Settings → Transcription — and decodes with a single vocabulary that spans all its languages, so mixed speech is not forced through a monolingual dictionary.
Language transitions. The multilingual vocabulary usually keeps short switches as spoken; long runs in the second language are the spans to review. For balanced two-language recordings leave detection automatic; if one language dominates, set it so a short opening in the other language cannot mislabel the file.
98 languages, on-device. The full catalogue is on /transcribe. Performance varies by language and by recording quality.
Translation and bilingual export. Translate a transcript on-device into 59 target languages; the original stays as the record. Documentary makers and researchers sharing tape with monolingual colleagues can export bilingual SRT/VTT — each original line paired with its translation.
Who benefits from multilingual transcription
International journalists and correspondents. Reporters covering global stories frequently conduct interviews in multiple languages — often within a single interview when a source is more comfortable in their native language but switches to English for technical or formal points.
Bilingual lawyers and consultants. Professionals serving immigrant communities often work with clients who code-switch between their heritage language and English during important moments. Capturing both languages provides a more complete record of the conversation.
Oral historians and ethnographers. Diaspora families switch languages mid-story; the transcript keeps the switch, which is often the finding.
Qualitative researchers studying language contact. Sociolinguists, anthropologists, and education researchers studying bilingualism, dialect contact, and language change produce recordings that are inherently multilingual. Kajo keeps both languages in order — the linguistic data that is the object of study.
Consultants and expert-interview professionals operating across linguistic boundaries. EU calls in mixed French and English, Latin American calls mixing Spanish and Portuguese, MENA work in Arabic and English — multilingual audio that requires capable transcription.
Language accuracy: honest expectations
Accuracy varies by language. English has the most training data and the lowest error rates; Spanish, French, German, Portuguese, Japanese, Korean, and Mandarin follow; Arabic, Hindi, Turkish, and many others land further down, and low-resource languages and heavily dialectal speech need the most review. The model’s developer publishes per-language error rates, linked from the engine guide; the free allowance lets you measure on your own recordings.
For languages with smaller training-data representation — certain regional languages, languages with a small internet footprint, or languages with significant dialectal variation — expect more post-transcription editing.
Practical tips for multilingual transcription
Record clearly. Microphone quality and recording environment affect transcription quality for all languages, but the effect is proportionally larger for languages where the model has less training data.
Specify the primary language when possible. If your content is predominantly one language with some code-switching, set the primary language in Settings → Transcription. If both languages are roughly equal, auto-detection is appropriate.
Review the switch points. Code-switching accuracy varies by speaker. Review the transcript sections where language transitions occur — these are most likely to contain errors.
Use the transcript as a draft. For all multilingual content, plan for a review phase. The AI generates a working draft; human review provides the accuracy that professional work requires.
FAQ
Can Kajo translate a transcript into another language? Transcription writes the spoken language as-is. Translation is a separate on-device step: Kajo’s built-in Gemma 4 model translates in-app into 59 languages, and bilingual subtitle export (SRT/VTT with original and translated lines) uses that path. No external translator or cloud service is involved.
How well does Kajo handle mid-sentence language switches? Best for language pairs that are well represented in the model’s training data — Spanish↔English, Arabic↔English, French↔English, and other pairs involving English. Short switches usually come through as spoken; long runs in the second language are the spans to review. Less common language pairs may be more variable — plan for a light review pass.
How does Kajo handle speakers with strong accents in a second language? Accented second-language speech is handled with reasonable accuracy for well-represented accent types. Heavy accents in less common language pairs may produce more errors. Clear pronunciation generally improves accuracy regardless of accent.
Conclusion: transcription for a multilingual world
The world is multilingual. Its professionals, researchers, and communities speak across language boundaries as a natural part of daily work and life. Transcription tools that demand everyone stay in one language are not just limited — they are exclusionary of genuine professional practice.
Kajo Voice handles languages as they are actually spoken: mixing, switching, and flowing between codes as speakers naturally do — on your machine, with the original kept as the record.
- Free — 5 files or 150 minutes of audio, total for life, full multilingual support
- Lifetime — one payment ($49 launch / $99 standard) unlocks unlimited imports and Lifetime convenience (watch folders, batch export); 98-language transcription, local and private
Related: For mid-conversation language switches (Spanish/English, Arabic/French, Hindi/English), see the Code-Switching Transcription Guide.