This is a copy. Use Snipmill at snipmill.com.

Skip to content

Turn a recording into text you can edit, copy or save as subtitles.

Settings

Settings

Runs on Snipmill's server: only the sound is sent, and deleted after. How we keep it private

Good to know

What Audio to Text does with your recording

Sound
What happens: Read once on your device and packed small. A video's picture is skipped, and a file with no sound is turned away
Upload
What happens: The sound alone, packed small: Opus, or MP3 where the browser can't make Opus
Words
What happens: Written on the server by a speech model
Repeats
What happens: Looping text is dropped: a phrase of 4 or more words said 3 times or more, or text that keeps repeating itself
Warnings
What happens: The bar warns when the language heard is unsure, isn't the one you chose, or is mixed; when the file sounds like music; or when the text comes out in another language's letters. Where it heard another language than the one you chose, "Use <language>" runs it again in that one
Lines and times
What happens: Each line carries the time the model gives it, so you can play it back. Voices are not told apart
Your file
What happens: Stays as it was; the text is new

Which setting to pick

Language spoken
Choice: Auto detect
Pick it when: One language throughout. The model names the language it hears at the start and writes the whole recording in that language; the list then shows the language heard
Language spoken
Choice: One of the 24 languages
Pick it when: The recording is short, starts with music or chatter, or detection named the wrong language. The language you pick is used for every part
Save as
Choice: TXT
Pick it when: Notes and documents: a new paragraph at each pause of 2 seconds or more. With "Times in copy, TXT, DOCX and PDF" on Include, one line each, like [1:07]
Save as
Choice: DOCX
Pick it when: Word, Google Docs and Pages: the file's name as the title, a line with the language, length and word count, then the text. Right-to-left languages are set right to left
Save as
Choice: PDF
Pick it when: To print or share: the same layout on A4, or Letter where your browser is set to the US, Canada or Mexico, with text you can select and search. Text in Latin, Greek or Cyrillic letters downloads as a PDF; other scripts open the print window, where you choose Save as PDF
Save as
Choice: SRT
Pick it when: Subtitles for YouTube and video players: numbered lines with start and end, like 00:01:07,450
Save as
Choice: VTT
Pick it when: Subtitles for web players and YouTube: starts with WEBVTT, times like 00:01:07.450

How accurate it is

Audio to Text uses a speech model on Snipmill's server, so it works in the browser on a computer, phone or tablet.

The page gives no accuracy percentage: it depends on the language, the recording and the speakers. A speech model can write words that were not spoken, can repeat itself, and does worse with some accents and dialects. Click a line's time to check it against the sound before you use it.

The 24 in the list are the languages it writes well: Bulgarian, Catalan, Chinese, Dutch, English, Finnish, French, German, Indonesian, Italian, Japanese, Korean, Malay, Norwegian, Polish, Portuguese, Romanian, Russian, Serbian, Slovak, Spanish, Turkish, Ukrainian and Vietnamese. We measured the model on Snipmill's server on recordings of read speech, 20 sentences a language: in these languages about one word in ten or fewer came out wrong, a little more in Korean and Serbian. How accurate it is on your file depends on the language and on how clear the recording is. Other languages aren't offered yet; on Auto detect, speech in one of those gives no text and the bar says "<Language> isn't offered: the text would be wrong."

Notes from a lecture

Saved as TXT, a lecture reads as paragraphs, a new one at each longer pause. Switch "Times in copy, TXT, DOCX and PDF" to Include to keep a time on every line, so a note can point back to the recording.

"Find in the text" shows only the lines with a term; click a line's time to hear what was said around it. Check terms and names against the slides, since a speech model can write words that were not spoken.

What people transcribe

  • A 45-minute lecture recorded on a phone: inside the 60-minute limit. Saved as TXT, it reads as paragraphs; "Find in the text" shows every line with a term.
  • A 90-minute meeting: over the limit, so cut it into two 45-minute parts with Audio Cutter and transcribe each. The times of the second part start again at 0:00.
  • A 3-minute talk in Spanish for subtitles: leave Auto detect and save SRT; each line keeps its times, like 00:01:07,450.

What goes to our server

Only the sound goes to Snipmill's server, when you press the button that starts the job; the line under it says so. The server deletes it once it has answered and keeps no copy. The page reads the recording and makes its sound small; a video's picture isn't sent. The text comes back to this page, and only you decide where it is saved.