Skip to content
VooHr Support Contact Sign in

Start here

Languages and transcription

Transcription and analysis language are separate choices. A language hint can guide speech recognition, while analysis language controls the language requested for the generated interpretation.

Speech recognition #

For audio, the transcription page sends a selected language hint with the file. The configured Deepgram speech-to-text provider maps supported hints to provider language/model settings; unknown hints fall back to English. The recognition result is text, not a guaranteed word-for-word record, so review names, code-switching, accents, specialist terms, and speaker changes.

The implementation includes hints such as English, English (India), Hindi, Hinglish, Spanish, French, German, Portuguese, Russian, Japanese, Korean, Chinese, Dutch, Italian, and automatic detection. Availability and quality depend on the configured provider/model and audio conditions.

  • Language hints include English, Hindi, Hinglish, Spanish, French, and German.
  • Other listed hints include Portuguese, Russian, Japanese, Korean, and Chinese.
  • Review names, accents, specialist terms, code-switching, and speaker changes.
  • For mixed-language speech, automatic detection is available but best-effort.

Analysis language #

Analysis has its own language selection. English is the default; additional options are drawn from the server's configured enabled-language list. The backend stores analysis results per language, so a later analysis in another language does not overwrite the earlier language entry. Some language instructions exist in code but are not enabled by default, and deployment configuration can affect the available list.

Choosing a different language requests another analysis; it is not simply a translation of the saved analysis. The model may express nuance differently across languages. Compare it with the transcript and use a fluent human reviewer when wording matters.

If a language looks wrong #

Check the selected transcription hint and make sure it matches the spoken language or mix. If the transcript is already poor, correct it in the editor before analysis rather than expecting the analysis model to repair recognition. For mixed-language speech, automatic detection is available as a hint but is best-effort, not a guarantee.

A language mismatch can alter names, negation, and the meaning of short phrases. Do not rely on a score or suggestion until you can verify the underlying text.

Use this guidance thoughtfully

This page describes the current product workflow; it is not an HR policy or a promise about every deployment. Before acting on information about an employee, verify the source record, consider context the system cannot see, and use your organization's process. If a control behaves differently from this guide, check the current workspace UI and confirm the saved result rather than assuming an action completed.