EasyWhisperUI: local Whisper transcription with a desktop UI and GPU checks

Transcription runs locally, and the app is offline after setup

EasyWhisperUI transcribes audio and video on your own hardware with Whisper and GPU acceleration, and it works offline once the required tools and models are downloaded.

That local-first property is the point of the app. Audio and video that never leave the machine is the reason to reach for a desktop Whisper wrapper at all, and the app is explicit that the offline behaviour is a feature rather than a limitation. The cost is that the first run does real work: fetching the models, and on macOS the bundled Whisper and FFmpeg arrive with the app itself.

The rewrite history explains the current shape. The author's note says the app has come a long way, from the original C++ app to a full rewrite in Electron, and now to a redesigned interface with more reliable setup and GPU checks. The repository layout reflects the same split, with an electron/ directory holding the current application, a legacy/ directory for the older code, resources/ for assets, and .gitmodules because the build needs submodules.

The primary language is TypeScript, and the project is a solo-maintained application rather than a company product. The author says donations are the project's only source of funding, and there is a public poll linked from the README for voting on which feature to work on next, which is an unusual and honest way to run a backlog.

Three platforms, and the Linux build is the one that is not ready

The download table in the README is the clearest statement of where the project actually stands.

Windows is listed for Windows 10/11 on x64, with a 3.1.0 installer as a Windows executable. macOS is listed for macOS 13 and later on Apple Silicon, with a 3.1.0 DMG. Linux is listed as Beta, tested on Ubuntu, with the 3.1 build pending and previous releases linked instead.

So two of the three platforms have a current binary and the third does not. If you are on Linux, you are running an older release, and the feature list you read may describe things your build does not have. That is the practical consequence of the beta label, and it is more useful information than the release notes give on their own.

The acceleration story is per platform. Windows uses Vulkan on supported GPUs and macOS uses Metal, and CPU-only processing is also available. The CPU-only path matters for two different reasons: a machine with no supported GPU, and a user who would rather not have a large model resident in video memory.

The release history is short and recent: v3.1 on 2026-10-09, v3.0 on 2026-10-08, and v2.1 earlier in 2026. The last push was on 2026-10-09, so the app is being worked on, though the repository records no licence that GitHub can classify because the licence is a custom one.

The quick start is four steps and two of them are about models

The setup flow is described in four steps. You run the Windows installer, or open the Mac DMG and drag EasyWhisperUI into Applications. You launch the app and let the initial setup finish, and on macOS that setup includes Whisper and FFmpeg. You choose your model, your input language and your output formats. Then you click Open or drop files into the window, transcription starts automatically, and exports are saved beside the original files.

The model choice has a rule worth knowing before you pick one. Use a multilingual model for non-English audio, because the .en models are English only. Missing models download automatically, and for a model you already have you choose custom and then Select Model File.

Translation into English is a separate step rather than a separate model. You choose a multilingual model such as medium, select the source language, and add --translate in the Arguments field. The Arguments field is also where the advanced controls live, since the feature list mentions extra Whisper arguments, a built-in CLI help and accessible setup logs.

Two smaller conveniences show up in the feature list and are the kind of thing that separates an app people keep from one they try once. Download progress shows size, speed and time remaining, and can be cancelled. And cached audio avoids repeated conversion when you retry or change models.

Batch queues, live transcription and the model download manager

The feature list is long enough to be worth taking apart, because the items are not all the same kind of thing.

Batch processing lets you drag in multiple files, skip individual items, or stop the queue. That is the mode you actually use for a folder of recordings, and the ability to skip and stop matters more than it sounds, since a wrong model on the first file otherwise means waiting for the whole queue.

Live transcription transcribes microphone audio in real time, and it is marked as beta. It is the feature most likely to be missing or rough on the Linux build.

Model management is a section of its own: you can download, manage, or load custom models, with color-coded memory estimates. The memory estimate is the practical bit, because it is what tells you whether a given model will fit alongside the rest of your applications on the machine you are on.

Output is multilingual transcription, translation into English, and TXT and SRT exports. The intended use of the output is spelled out too: give the TXT file to a language model for summaries, study notes or a cleaned-up transcript, and export SRT to add subtitles in a video editor or player.

The interface itself offers Light, Dark and Auto themes, with translucent backgrounds and a spacious output console, and setup and repair is a guided installation with a clean reinstall path on Windows and Linux that preserves your models and settings.

Troubleshooting on macOS means telling macOS to let it run

The troubleshooting section is short, and the macOS part is the one that will hit you first.

The Mac app is ad-hoc signed and not notarized, so macOS may block it. The instruction is to use System Settings, Privacy and Security, and Open Anyway. If you get a damaged message instead, the README points at the release's own first-opening instructions.

That is the practical consequence of shipping an unsigned build outside the App Store, and it is a real friction point for every new macOS user. It is also the reason the download table lists macOS 13 and later: the Gatekeeper behaviour described applies to current systems.

On Windows and Linux the recovery path is Settings and Clean reinstall, and on Mac you restart the app. If the problem continues, Settings and Show log file opens a log you can attach to an issue, and the issue template asks for the error, the log, the operating system, CPU, GPU, RAM, app version and model. If there is no setup log, you copy the console error instead.

That list is a good summary of what actually varies between machines, since the hardware fields are exactly the ones that decide which acceleration path you get and which model fits in memory.

A proprietary licence over open-source components

The licence is the constraint to read before anything else, and it is stated as a block of text in the README rather than as a recognised open source licence:

Copyright (c) 2026 Mehtab Mahir
All rights reserved.

This software is proprietary and the following is not allowed for commercial purposes:
it may not be copied, modified, distributed, or used without explicit permission from the author.

Those actions are permitted for personal use ONLY.

So the application itself is proprietary, and personal use is the only permitted case. The licence then lists the open-source components included in the distribution, starting with whisper.cpp by Georgi Gerganov under the MIT licence, and FFmpeg, whose licence depends on the build, with the note that the bundled macOS binary is GPL 3.0 or later and that the Windows setup downloads a GPL build from BtbN.

That FFmpeg detail is the one to sit with. A proprietary application that bundles a GPL 3 binary is a licensing question for the author, not for you as a user, and it is also the reason the Windows build downloads its FFmpeg rather than shipping it. The credits list names the rest of the stack: SDL2, electron-builder, and the macOS FFmpeg builds.

For a user, the practical effect is simple. Install it, use it for your own transcripts, and if the work you are doing is commercial or you intend to redistribute the app, you need permission from the author rather than a licence grant.

Editorial conclusion

Adopt EasyWhisperUI if you want Whisper transcription on your own hardware with a desktop interface, you want the model and the audio to stay on your machine, and personal use is what you need it for. Do not adopt it for anything commercial, because the licence in the repository states the software is proprietary and that copying, modification, distribution or use without explicit permission is not allowed, with personal use the only permitted case. Before you install, check that your platform has a build: the Windows installer and the macOS DMG are published for version 3.1.0, while the Linux build is listed as pending and beta.

Frequently asked questions

Official sources

1. Issues
2. mehtabmahir/easy-whisper-ui on GitHub
3. README
4. Releases