Transcribing Multilingual Meetings: What to Know Before You Build

The multilingual meeting problem

Global teams have meetings in mixed languages — English with French terms, a Chinese sales call that switches to English for technical terms, a Spanish customer support call that references English product names. Naive transcription pipelines trained on monolingual data fail on these.

How language detection works

Modern transcription engines (Whisper-based pipelines included) can detect the language automatically from the first few seconds of speech. They use this to select the right acoustic model and language model for the rest of the call.

The challenge is code-switching — when speakers change languages mid-sentence or mid-call. Most engines handle this by running multi-language models that can handle switching at the sentence boundary, but accuracy drops compared to a purely monolingual call.

Specifying a language explicitly

If your users primarily speak one language, specifying it explicitly improves accuracy:

POST /v1/audio/transcriptions
{
  "language": "fr",   // ISO 639-1 code
  "file": <audio>
}

Supported codes include en, fr, de, es, ja, zh, pt, it, ko, ar, and many others.

What to tell your users

Be upfront about accuracy expectations for non-English calls. A French call transcribed with the default English model will be garbled. A French call with language: "fr" set will be accurate. Give your users a way to set their preferred language in your product.

Speaker labels across languages

Speaker diarization works independently of language — it identifies speakers by voice characteristics, not by what they say. So even on a mixed-language call, the speaker attribution remains accurate even when the transcribed text has errors.