By Andrew Armstrong 
If you have ever tried to get a transcript out of an hour-long recording without paying for it, you will know the shape of the problem. Otter gives you 300 minutes a month on the free tier with a 30-minute cap per recording, and the paid tiers run $17 to $24 a month. Adobe’s Enhance Speech is cloud only, so the audio has to leave your machine. And OpenAI’s Whisper — which is free, unlimited and very good — wants Python, a command line and a manual model download before it will do anything at all.
Mac users have had MacWhisper for this for years. Windows users have had a choice between a subscription and a terminal. TranscribeGeek is the third option: a proper Windows application that runs Whisper locally, with nothing to sign up for and no limit on how much you transcribe.
What it does
Drop audio or video files in, pick a model and a language, and press Transcribe. Each file gets a plain text transcript written beside it, and a SubRip .srt subtitle file too if you want one.
- Audio and video — WAV, MP3, M4A, FLAC, OGG, Opus, WMA, AAC, MP4, MKV, MOV, AVI and WebM.
- Timestamps, optionally — the text file can carry a timestamp per segment or be clean prose.
- Subtitles — a standards-correct
.srt, written with the invariant culture so players do not choke on a decimal comma. - A real queue — drop in twenty files and walk away.
- 22 languages, or automatic detection.
The queue is worked one file at a time on purpose. Whisper already uses every core it can find, so running two files at once makes both of them slower and the progress bar meaningless.
What it will not do
Every app in the range has a Settings screen that states plainly what it refuses to do. TranscribeGeek’s says this:
- It does not send your recordings anywhere. Transcription runs in the application, on your own processor. There is no account and no server. The only thing the app ever downloads is a speech model, from the Models screen, when you ask for one.
- It does not overwrite your files. If
interview.txtis already there, the next one isinterview (2).txt. Quietly replacing somebody’s earlier transcript because they ran the file twice is not acceptable behaviour for a tool that is supposed to be safe. - It does not bundle ffmpeg. More on that below.
- It does not count anything. No per-minute cap, no monthly allowance, no trial.
Why there is no speech model in the installer
Whisper comes in sizes, and they are not small. Bundling them all would make the download several times bigger for no benefit, because most people only ever use one. So TranscribeGeek ships with none of them and fetches the one you pick, once, with the size shown before anything starts.
- Tiny — 78 MB. Fastest, least accurate. Good for checking that a file transcribes at all.
- Base — 148 MB. A sensible starting point on an older machine.
- Small — 488 MB. The best balance for most people. Start here.
- Medium — 1.5 GB. Noticeably better on accents and poor recordings, and several times slower.
Models are kept under your own profile rather than beside the program, so a reinstall does not mean downloading a gigabyte again. There is a Remove button on the Models screen when you want the space back.
About ffmpeg
MP3, MP4 and everything else that is not a plain 16 kHz mono WAV has to be decoded before Whisper can read it, and ffmpeg is what does that. TranscribeGeek runs ffmpeg as a separate program and never links it — that is a licensing decision, not a style one: the commonly distributed builds are GPL, and linking one into the application would impose terms we have not signed up to.
It is not bundled either. If ffmpeg is not on your machine the app says so plainly on the Transcribe screen and still handles WAV, rather than silently pulling down a 90 MB binary you did not ask for. Put ffmpeg.exe next to TranscribeGeek or anywhere on your PATH and everything else starts working.
How fast is it?
On an ordinary desktop processor, roughly real time with the Small model — an hour of audio takes about an hour. Tiny is several times faster and less accurate. Medium is several times slower and better on difficult recordings. There is no GPU build, because the CUDA runtime would add several hundred megabytes that most people cannot use.
Requirements
- Windows 10 version 1809 or later, 64-bit. Windows 11 fully supported.
- Runs as a normal user — no administrator rights are needed or requested.
- About 71 MB for the installer. .NET 8 is inside it; there is nothing to install first.
- GPL-3.0. Free to use, including at work. No paid tier, ever.
Where to get it
The installer and a portable build are on the TranscribeGeek product page, with SHA-256 checksums for both. The source is on GitHub.
Speaker labels — who said what — are the next thing planned. If you hit a bug or want something adding, open an issue and it will be read.
Discover more from TechyGeeksHome
Subscribe to get the latest posts sent to your email.
