VoxScribe

Offline, open source dictation for your PC and your phone.

Free, open-source voice typing for Windows and Android. Talk, and your words are typed straight into whatever app you are using. Your speech is transcribed locally on your own device and never uploaded anywhere.

• Windows: hold a hotkey (F9 by default) anywhere, talk, release. Also turns audio and video files into text and .srt captions. Download the installer or run winget install ahmedhmam1994.VoxScribe.
• Android: a small floating button appears above your keyboard in any app. Hold or tap it and talk; it works in airplane mode. Get the APK.

No account, no subscription. MIT licensed.

Screenshots

Caption a whole folder in one run. Drop in a folder of episodes and VoxScribe writes an .srt next to each file. It never overwrites an existing caption file, reports what it saved, skipped and failed, and flags cards that look too fast to read or follow a long silence. Failed files can be retried with one click.

Fix the words the model keeps getting wrong. Add a short vocabulary list for names and jargon, and word replacements for anything it mishears the same way every time.

Features

New here? The full tutorial walks through every feature step by step.

What Pro adds: the optional one-time $14 unlock covers Snippets, fill-in templates, per-app profiles (including per-app term packs), dictation modes that switch by app, Code mode and code dictation, smart formatting, a second language key, the Medium, Large turbo and Large Whisper models (also used by Try harder), folders of files beyond 2 free runs, .vtt, .md and .docx export, and speaker labels. Everything else in this list is free.

• Hold-to-talk hotkey (F9 by default, changeable): works system-wide, in any app; change it from the tray menu's "Change Hotkey..."
• Direct typing into the focused window: not clipboard/paste, so it never overwrites what you've copied
• 100% local transcription: powered by faster-whisper running on your CPU (or your NVIDIA graphics card, optional); your audio never leaves your machine
• More speech models, only if you want them: Cohere Transcribe (strong Arabic that keeps your dialect), Parakeet (fast and accurate English plus 24 European languages) and Moonshine Arabic, each downloaded only when you pick it in Settings, Speech. Whisper stays the default and takes over if a model fails
• Start meeting: records your microphone (You) and what the PC plays (Others), removes the echo when you use speakers, and writes a markdown note with action items, decisions, questions and the transcript into a folder you choose, for example your Obsidian vault. Everything stays on the PC. VoxScribe asks you to confirm before it records, and you should tell the other people on the call that it is being transcribed
• Term packs for medicine, law and business: words like MRI, metformin, habeas corpus and EBITDA are written correctly with any speech model. Import your own list as a plain text file. With Pro, pick packs per program
• Teach VoxScribe your voice: read a short script and it finds the words it mishears in your voice or accent, then offers to fix them from then on
• Spoken formats: "one twenty over eighty" becomes 120/80, "five milligrams twice daily" becomes 5 mg BID, "section twelve subsection b" becomes section 12(b), "fifty thousand dollars" becomes $50,000. Off by default
• Notes inbox key: hold it and speak, and the words are saved as a timestamped line in a markdown file instead of being typed
• Read it back and unsure words: hear each dictation in Windows' own voice, and see the words the model was unsure about underlined in the transcript. Both are off until you turn them on
• Rule-based text cleanup: strips filler words ("um", "uh", "like", "you know") without calling any paid AI API
• Spoken commands: say "period", "new line" or "question mark" for symbols, "scratch that" or "delete last word" to take back what you just typed, and "capitalize that", "all caps" or "replace X with Y" to change it. An optional "send it" presses Enter
• Custom vocabulary and word replacements: teach it names and jargon, and fix words it mishears the same way every time. Right-click a word in the transcript to always correct it, and export or import both lists as one file (the Android app reads it too)

Download

Grab the latest installer from the Releases page.

Prefer a package manager? With winget, which ships with Windows 10 and 11:

winget install ahmedhmam1994.VoxScribe

Or with Scoop:

scoop bucket add voxscribe https://github.com/ahmedhmam1994/scoop-voxscribe
scoop install voxscribe

Note on Windows SmartScreen: since VoxScribe is a new, unsigned app that hooks the keyboard (for the hotkey and to type text), Windows may show a "Windows protected your PC" warning on first run. Click More info → Run anyway to proceed. This is expected for unsigned indie software and not a sign of a problem.

The installer does not require admin rights (per-user install). On first launch, VoxScribe downloads the Whisper speech model (a few hundred MB) from Hugging Face. This requires an internet connection once; after that, transcription works fully offline.

Android

VoxScribe also has an Android app. A small floating button appears above your keyboard whenever you tap a text box, in any app. Hold it, talk, release, and the text is typed into the box. It works with whichever keyboard your phone already uses, and your voice is transcribed on the phone, never uploaded.

Install: download VoxScribe-Android-1.1.2.apk from the Android 1.1.2 release and open it. Android will ask you to allow installs from your browser or file manager. It is an APK only for now, not on the Play Store.

Automatic updates: add this repository to Obtainium (Add App, paste the repo URL) and set "Filter release titles by regular expression" to Android, so it skips the Windows releases.

• Android 8 or newer, 64-bit phone (almost every phone from the last several years)
• The speech model is bundled in the app, so the download is about 115 MB and it works in airplane mode
• Setup takes three permissions, and the app walks you through them: microphone, the Accessibility service (Android requires it to let an app type into other apps), and display over other apps
• Tap the button to start and stop, or hold it. Drag it to the bottom of the screen to hide it for 10 minutes. It never appears in password or PIN boxes
• Transcribe audio and video files on the phone and save the text or .srt captions. A whole folder at once has 2 free runs, then it is Pro
• Dictation modes (Chat, Email, Code, Notes), switching by app with Pro
• Email, web address and number boxes get the text the way they expect, for example "sam at gmail dot com" becomes [email protected]
• Spoken commands ("period", "new line", "scratch that", "capitalize that", "replace X with Y"), an optional "send it" that taps the Send button, "Fix a word", and word replacements with export and import are free, each with its own page in Settings. Smart formatting, code dictation, fill-in templates and per-app profiles unlock with the same Pro key as Windows, which also unlocks Snippets
• When a newer Android release is published, the app shows an update banner with a download button

The Android source is in android/voxscribe_app: Flutter for the app screens, Kotlin for the floating button, typing service and speech engine.

Requirements

• Windows 10 or 11
• A working microphone
• ~1GB free disk space (app + downloaded model); the optional speech models and the NVIDIA GPU option add their own downloads, which you choose

How it works

1. Hold F9 (or your chosen hotkey, see the tray menu's "Change Hotkey...")
2. Speak
3. Release the hotkey. The transcribed, cleaned-up text is typed into whatever app you're focused in

Privacy

• Audio is captured, transcribed, and discarded locally: your audio and text are never sent to a server
• Text cleanup is regex-based, running entirely on your machine: no AI API calls
• Network use is limited to: the one-time Whisper model download on first run (plus any optional speech model or GPU files you choose to download), a check for new versions on GitHub, anonymous usage pings (event name, app version, OS version and a random id; on launch (every launch on Windows, at most once a day on Android), the first time you dictate, and once on each day you dictate; can be turned off in Settings), and a one-time activation request if you buy Pro. See the privacy policy for details

Building from source

python -m venv venv
venv\Scripts\pip install -r requirements.txt
venv\Scripts\python.exe main.py

See CLAUDE.md for architecture notes and build/packaging commands (PyInstaller + Inno Setup).

Command line

The same local engine can be called from scripts. From a source checkout:

venv\Scripts\python.exe cli.py meeting.mp3
venv\Scripts\python.exe cli.py talk.mp4 --format srt --output captions
venv\Scripts\python.exe cli.py interview.m4a --language auto --translate

Text goes to stdout, or into the --output folder (an existing file is never replaced). Plain text is cleaned and uses your word replacements, like dictation; --raw gives it exactly as heard. Plain text and .srt are free; .vtt, .md and the three most accurate models need Pro, as in the app. Run cli.py --help for every option.

Code signing

The installer is currently unsigned, so Windows SmartScreen shows an "unknown publisher" warning on first run. VoxScribe is applying for a free certificate through the SignPath Foundation's open-source program: see CODE_SIGNING.md for the policy that program requires. Paid alternatives also work: a traditional OV/EV certificate from a CA (SSL.com, Sectigo, roughly $70 to $400/yr), or Microsoft Trusted Signing via Azure (usage-based, no hardware token required). With any of these, run scripts\sign_release.ps1 after building; see that script's header comment for exact usage.

License

MIT