Transcribe audio recordings to text, with your own vocabulary

You already make recordings: voice memos between appointments, meetings, interviews. All that remains is to turn those files into text you can use. TechVoice transcribes them in moments and corrects the terms of your profession using your own vocabulary, just as it does for live dictation.

How it works

1. Upload the file to a note, from your computer or your phone.
2. TechVoice transcribes it with Voxtral Small Transcribe 3, Mistral AI’s speech-to-text model for audio files, set to the language of the app, French or English.
3. Correction is applied using your vocabulary and your past corrections, without rewording anything.
4. You review it: the text lands in the note, the corrections are visible, and one-click formatting turns it into an email or polished prose if needed.

Fast, even for an hour of audio

The model receives the whole recording, so it has all the context before it writes anything, which makes it more accurate than live dictation. On a real 30-minute meeting, transcription took between 52 and 67 seconds in our measurements, a little longer at peak times.

The limit is 60 minutes and 100 MB per file. It is not a TechVoice restriction: it is how much the model can process in one go, as the audio and the text it produces share its capacity. We explain these constraints in our hands-on review of Voxtral.

Your specialist terms, spelt right

A recording of a meeting or a consultation is full of specialist vocabulary, and that is exactly where general-purpose transcription fails. On our orthopaedic test corpus, in French, the raw transcript of an uploaded file got 9.6% of specialist terms wrong, and 2.6% after correction with the practitioner’s vocabulary. These figures come from synthetic voices and are optimistic; above all, they show what correction adds.

This is the same vocabulary as for dictation: your field, your sample documents and every one of your corrections. A correction made on a dictation also benefits later transcriptions, and vice versa.

What transcription does not do

• It does not separate speakers: the text is continuous, with no speaker names or timestamps.
• It does not summarise: you get what was said, corrected, not a summary. One-click formatting can then turn it into an email or polished prose for you to review.
• It does not replace the consent of the people recorded: only record with the agreement of those present.
• It is not for identifiable health data: TechVoice does not hold HDS certification, France’s certification for hosting health data; do not upload a recording in which a patient can be identified.

Formats and privacy

Common formats are supported: m4a (iPhone voice memos), mp3, wav, aac, ogg, opus, flac, webm and mp4. The file is transcribed by Mistral AI, in France, then TechVoice does not keep it; only the text remains, in your notes. Full details are on our Privacy page (in French).