Turn a recording into text
Published · Updated 3 min read
Transcribing a recording, a lecture, a meeting, a voice message or an interview saves you from listening to everything again to find a sentence. The tool transcribes an audio file in your browser, with a speech recognition model. The result is useful, but it needs a reread: on a clear voice, expect about one word in four to correct, and the choice of language as well as noise change the result a lot.
What the tool does
Audio to text transcribes an audio file in your browser. The file is not sent: only the model is downloaded, once, after you tick the consent box. Three export formats are offered: TXT for a text, SRT and VTT for subtitles.
What the first use downloads
On first use, the browser downloads the model and its scripts from two services outside the site. These are downloads: none carries your audio file. The "Clear cache" button then removes the model from the browser if you want to get the space back.
Choose the language
The language menu offers an automatic setting and twenty languages. With the built-in model, automatic detection is easily fooled: on a French voice, it may take it for English and produce sentences unrelated to what was said. The tool therefore selects the page's language by default. Always choose the language of the recording, it is the setting that changes the result the most.
| Language set | Wrong words out of 32 |
|---|---|
| Automatic | almost all |
| French | 8 to 10, a quarter |
What a correct language gives
On a clear voice, the transcription is understandable, with typical errors: a lost plural ("parcels" for "parcelles"), a proper name without a capital, a number written in digits ("3" for "three") and rare words misheard. That is why rereading remains necessary.
The effect of noise
Background noise quickly degrades the transcription.
| Recording | Wrong words out of 32 |
|---|---|
| No noise | 8 (25 %) |
| Light noise, weaker than the voice | 11 (34 %) |
| Noise as loud as the voice | 31 (97 %) |
Light noise costs a few words. Noise as loud as the voice makes the transcription unusable.
Processing time
Expect a few seconds to about ten seconds for a recording of about fifteen seconds, model loading included, on a desktop computer. A longer file needs more time and memory.
What these figures do not say
These orders of magnitude come from clear, regular voices. A real voice, with an accent, hesitations or several people speaking, will give other figures. Try a short extract of your own recording before running the whole file.
Improve the result
- Choose the language of the recording in the menu.
- Reduce the noise at the source: a close microphone gives a better result than a filter applied afterwards.
- If the file is in an unusual format, convert it first with Convert a voice recording without damaging it.
Always reread proper names, numbers and technical terms.
For which recordings
A lecture, a meeting, an interview or a voice message: the text lets you search for a sentence without listening to everything again and take passages from it.
Check the resulting text
- Listen to the recording while following the text.
- Correct the names, the numbers and the agreements.
- Export as TXT for a text, as SRT or VTT for subtitles.
Frequently asked questions
- Is my recording sent to be transcribed?
- No. Only the model is downloaded, once, after your consent. The audio file stays on your device.
- What accuracy should you expect?
- On a clear voice in French, expect about one word in four to correct. Background noise degrades it quickly.
- Which language should you choose?
- That of the recording. Automatic detection is easily fooled with the built-in model.
- Which export formats are offered?
- TXT for a text, SRT and VTT for subtitles.