Changelog

What changed, and when.

Product changes with the date they went live. The site launched on 14 September 2026.

2026-09-25

Small model for phones, CPU fallback when the graphics backend refuses the model

An iPhone report showed the download completing and the tab crashing as transcription began, then "a problem repeatedly occurred". Phones and other low-memory devices now run in phone mode: the tiny English model (about 120 MB) on the processor instead of the graphics card. The picker still offers the other models.
Diagnostic address switches (?device=, ?decode=, ?model=) documented on the browser support page, so other combinations can be tried from a phone without a new release.
If the graphics backend fails to load the model, the tool retries once on the processor before reporting an error.
Verified on an iPhone 17 in Safari and Chrome the same day. Android phones are still untested.

2026-09-25

Microphone mode: press "Use the microphone" and the words appear a few seconds after you say them, transcribed on your machine. The audio is kept in the tab so it can be played back and exported afterwards; nothing is sent anywhere. (The first deployment shipped with a security header that denied the microphone to the page itself; fixed the same day.)
The result view plays the audio in sync with the transcript: click a timestamp to play from there, and the current segment is marked as it plays.
Edit mode: correct text and timestamps in place. Exports use the corrected transcript.
Find and replace across the whole transcript, and a timing shift that moves every timestamp by a number of seconds.
Model picker in the drop card: the tested English base model, the multilingual base model with a language list, and the large-v3-turbo model. The two new options are marked untested until we have run them ourselves.
Storage line on the home page shows how much the site holds on the device and offers to clear the cached model files.
New page: subtitle tools, for converting between SRT, VTT and plain text and fixing timing offset or drift, all in the browser.
Benchmarks page with measured runs on named hardware, including 5 and 25 minute files, noisy, quiet and fast variants of the sample, and CPU mode.

2026-09-25

Browsers without WebGPU can now run the speech model on the processor through WebAssembly. It is single-threaded and slower than WebGPU (46 s for the 71 s sample on our test laptop, download included); the page says so before you commit.
Browsers without WebCodecs decode the whole file with Web Audio instead of streaming it. Works for shorter recordings; very long ones can run out of memory, which the notice explains.
The drop card shows what your browser can do before you pick a file, including the graphics adapter when the browser names it.
A "Try it on a sample" button runs the transcriber on a bundled 71-second recording, so you can see it work without a file of your own.
Result view reports the backend used and the real-time factor of the run.
Ads are no longer shown inside the transcript or under the drop zone, and not at all when the tool cannot run in the visitor's browser.
New pages: examples, browser support, editorial policy, author pages and this changelog.

2026-09-15

Decoding and transcription are interleaved, so memory stays flat at about a minute of audio however long the file is.
Fixed the window cut that stalled transcription near 4% on files longer than about two minutes.
Lighthouse pass: fonts self-hosted, stylesheets inlined, the ad script deferred until after load.

2026-09-14

FreeTranscribe goes live at free-transcribe.com: in-browser transcription with the base English Whisper model on WebGPU, streaming decode through WebCodecs and Mediabunny, live transcript with search, and TXT, SRT and VTT exports.
AVI files with MP3, PCM or AAC audio are demuxed by our own RIFF reader; WMV, FLV and MPEG-PS are refused with a conversion hint.
Silence skipping and repeated-phrase filtering to reduce the model's habit of inventing text over quiet passages.