Transcribe an .aup4 or .aup3 Audacity project to text and SRT
Drop a project, pick the language that is spoken, and get the words back as plain text or as .srt subtitles with times. The speech recognition runs on this device; the recording never leaves it.
An Audacity 1.x or 2.x project (.aup)? — the .aup and its _data folder together, or drop that folder or a .zip of it above.
Several projects? Drop or choose them together: export them all to one ZIP at once (Pro), or open them one by one free.
How do I turn an .aup4 file into text?
An .aup4 is an Audacity 4 project, an SQLite database of audio blocks rather than an audio file, so a transcription site cannot read it. Either export the mix to MP3 first and transcribe that, or drop the .aup4 on this page: it mixes the project's tracks and runs Whisper speech recognition in your browser, giving plain text and .srt subtitles with timestamps. The audio is never uploaded. Inside Audacity itself, Intel's OpenVINO plugins add Whisper transcription to Audacity 3.7.1 on Windows, writing the text into a label track.
Drop the .aup4 or .aup3 file (or an old .aup with its _data folder). Nothing is uploaded.
Play the mix to check it is the right recording.
Pick the spoken language and press Transcribe — the first time, the model downloads from this site.
Read the transcript, then download it as .txt or as .srt subtitles. Transcribing is part of Pro.
AUP4 & AUP3 to Text Converter — without uploading your audio
Transcription sites take audio files, and a project is not one: upload an .aup4 and they report an unsupported format. The usual way round is two steps — export the mix, then upload the MP3 to a transcription service — and the second step sends the recording to someone else's server. Here both happen in the browser.
The project's tracks are mixed down a few minutes at a time, and each piece is handed to Whisper, the open speech-recognition model, running in a worker on this page. The model is downloaded from this site the first time and kept by the browser. Transcribing is part of Pro; exporting the mix as MP3 or WAV stays free up to 256 MB, if you would rather use another transcriber.
How the transcript is made, and where it can go wrong
The project is mixed to one channel at its own sample rate, five minutes at a time, by the same mixer that plays it and exports it, so clip positions, envelopes, track volume and mute are applied as you hear them. The browser's own audio engine resamples each piece to 16 kHz, the rate Whisper listens at.
Whisper reads audio in 30-second windows. Each five-minute piece is cut into windows that overlap by five seconds on each side, and the overlaps are used to stitch the text back together, so a word on a window's edge is not lost twice. Each line of the transcript keeps the time Whisper gives it, shifted to where its piece starts in the project; that is what the .srt is built from.
The model is Whisper base, released by OpenAI, in the ONNX export published as onnx-community/whisper-base with 8-bit weights: an encoder of 23 MB and a decoder of 54 MB. The site serves the files of one fixed revision, checked against their recorded hashes when the site is built, and the ONNX runtime's WebAssembly build (14 MB) beside them. Nothing about the recording, the transcript or the file name is sent anywhere; only whether a transcription started, finished or failed, its language, and how many minutes it took are counted.
- Input
- .aup4, .aup3, and .aup with its _data folder
- Output
- Plain text (.txt) and SubRip subtitles (.srt)
- Model
- Whisper base (74 M parameters), ONNX, 8-bit weights
- First download
- About 96 MB from this site, then kept by the browser
- Runs on
- This device: WebAssembly in a Web Worker, one thread
- Measured
- Apple M3, Chrome: 34 s of English in 10 s, 20 s of Chinese in 5 s
- Language
- Chosen by you (20 offered); Chinese as simplified or traditional
- Tier
- Pro; the mix as MP3/WAV stays free up to 256 MB
New to the format? What an .aup4 file is, table by table — and how it differs from .aup3.
Sources
Where the facts on this page were checked. Audacity's code is cited at the tagged release it was read from.
- openai/whisper README — Whisper's model sizes: base has 74 M parameters and is multilingual; it was trained for speech recognition across many languages.
- onnx-community/whisper-base (the revision served here) — encoder_model_quantized.onnx (23,201,314 bytes) and decoder_model_merged_quantized.onnx (53,693,315 bytes), the 8-bit weights this page loads.
- intel/openvino-plugins-ai-audacity README — The Whisper Transcription feature generates a label track with the transcription; installation packages are published for Windows (latest v3.7.1-R4.2, 20 December 2024), with build instructions for Linux.
Converting this file
Q. Is my recording uploaded to transcribe it?
No. The project is read and mixed on your device, and the speech model runs on your device too, in a Web Worker using WebAssembly. What comes over the network is the model itself, once: about 96 MB from this site, which the browser then keeps. Nothing is fetched from Hugging Face or any other server while a project is open.
Q. Which languages does it understand?
Whisper was trained on dozens of languages; the list here offers twenty of them, from English, Chinese and Spanish to Russian, Japanese and Arabic. Pick the language that is spoken — it is not guessed. Whisper writes Chinese in traditional characters, so choose "Chinese, simplified" to have the text converted to simplified characters with OpenCC, or "Chinese, traditional" to keep them.
Q. How long does it take, and how accurate is it?
It runs on the processor, in one thread. On an Apple M3 in Chrome, 34 seconds of spoken English took 10 seconds and 20 seconds of Chinese took 5, with the model already downloaded; a phone is several times slower, and a long recording takes a long time. The model is Whisper base (74 million parameters), small enough to download into a browser: clear speech comes out well, but expect mistakes with names, jargon, music or several people at once — in the Chinese test it heard 播客 (podcast) as 播可. Read the transcript before you publish it.
Q. What can I do with the .srt file?
An .srt (SubRip) file is numbered captions, each with a start and end time to the millisecond. YouTube, VLC and most video editors read it, so it can go straight onto the video made from the same recording — the MP4 export on the AUP3 to MP4 page, for instance. The times come from Whisper, a few seconds per line, measured from the start of the project.
Q. Is there a free way to transcribe an Audacity project?
Yes. Export the mix as MP3 or WAV (free here for projects up to 256 MB) and give that file to any transcription tool, including Whisper run on your own computer. On Windows, Intel's free OpenVINO AI plugins for Audacity 3.7.1 include a Whisper transcription analyzer that writes the text into a label track inside Audacity. What Pro adds here is doing it straight from the project, on any device, with nothing installed.
Free to open, play and export the stereo mix for projects up to 256 MB; larger projects still open and play free, and downgrading to .aup3 is free up to the same size. Recovery, per-track stems, and exports or downgrades above that size are a one-time $9.90 license — no subscription, no recurring fees.
Every job an Audacity project needs
Same local-first engine, one page per task.