Sorify turns audio and video files into text on your own PC. Drop files on the window, press Start, and get clean paragraphs, timestamped text or subtitles. It is built on OpenAI Whisper, runs on your NVIDIA GPU when there is one and on the CPU otherwise, and never uploads your recordings.

⬇ Download the latest installer  ·  Windows 10 / 11, 64-bit  ·  free for personal and commercial use

Highlights

• Drag and drop files or whole folders (sub-folders included). 30+ formats: mp3, wav, m4a, aac, flac, ogg, opus, wma, mp4, mkv, mov, webm, avi, ts and more. Video files use their audio track.
• Live transcript appears while the file is being decoded, as paragraphs or with timestamps.
• Every Whisper model, from tiny to large-v3. Auto picks turbo on a GPU and base on the CPU; the panel shows each model's download size, speed and GPU memory need.
• 99 languages, detected per file or chosen by hand, plus Translate to English.
• Outputs: text paragraphs, timestamped text, SRT, VTT, TSV and full JSON (segments and words). Save next to each file or into one folder; existing transcripts are skipped unless you choose Overwrite.
• Pro options: beam search or fast greedy decoding, word-level timestamps, subtitle line width (e.g. 42 for two-line captions), a vocabulary prompt for names and jargon, a switch for repetition loops.
• Right-click → Transcribe with Sorify in Explorer, and a sorify command for batch jobs from any terminal.
• Private by design: after setup it works offline. Files and transcripts stay on your disk.
• Self-updating: checks GitHub Releases at launch and asks before installing (Update now · Remind me later · Skip this version).

Install

1. Download Sorify-Setup-x.y.z.exe from the latest release.
2. Run it. It installs for the current user only (no administrator rights) into %LocalAppData%\Programs\Sorify.
3. Pick the extras you want: desktop shortcut, Explorer right-click menu, sorify on PATH.
4. After the files are copied, Setup downloads the components that are too big to ship: PyTorch (about 2 GB for the GPU build, 120 MB for the CPU build), Whisper and ffmpeg (80 MB). A small window shows the progress.
5. The first transcription downloads the speech model you picked (turbo 1.5 GB, base 140 MB). Models are kept in %UserProfile%\.cache\whisper and are shared with any other Whisper tool on the PC.

"Windows protected your PC" — the installer is not code-signed yet (certificates are expensive for a free app), so SmartScreen shows this warning the first time. Click More info → Run anyway. You can verify the download first: every release lists its SHA-256 in SHA256SUMS.txt, and certutil -hashfile Sorify-Setup-x.y.z.exe SHA256 prints yours.

Requirements

	Minimum	Recommended
Windows	10 (1809) or 11, 64-bit	11
Disk space	2 GB (CPU build)	6 GB (GPU build + turbo model)
Memory	4 GB	8 GB
GPU	none (CPU mode)	NVIDIA GeForce 10-series or newer, driver 560+ (CUDA 12.6), 6 GB VRAM for turbo
Network	for setup, model downloads and updates

Without an NVIDIA GPU Sorify uses the CPU: base runs faster than real time on a modern CPU, small is a good balance, the large models are slow.

Using Sorify

Queue. Drop files or folders anywhere on the window, or use Add files / Add folder. Each row shows length, status, detected language and time taken. Right-click a row for Open transcript, Show in folder, Transcribe again and Remove. Double-click opens the transcript.

Model. Auto is right for most people. Use medium or large-v3 when you need Translate to English (turbo does not translate). The .en models are English-only and a bit better at it on the small sizes.

Language. Auto-detection looks at the first 30 seconds of each file. Set the language when a short or accented recording is guessed wrong.

Save to. Next to each file writes talk.txt beside talk.mp3; One folder collects everything in one place (files with the same name from different folders get the folder name in front).

Advanced. Accurate uses beam search; Fast is greedy and roughly twice as quick. Word-level timestamps makes JSON word timings and better subtitle cuts. Subtitle line width re-cuts SRT/VTT captions. The vocabulary prompt tells Whisper how names and jargon are spelled. Switch off Carry context between chunks if a transcript starts repeating itself.

Keyboard. Ctrl+O add files · Ctrl+Enter start · Esc stop · Delete remove selected rows · Ctrl+A select all · F1 about.

Explorer. With the right-click option installed, select one or more audio/video files, right-click (on Windows 11: Show more options) and choose Transcribe with Sorify. The files land in the queue of the open window.

Models

Model	Download	GPU memory	Speed	When
tiny, tiny.en	72 MB	1 GB	~10×	quick rough drafts
base, base.en	139 MB	1 GB	~7×	the CPU default
small, small.en	461 MB	2 GB	~4×	good balance on CPU
medium, medium.en	1.5 GB	5 GB	~2×	accurate, good translation
large-v2, large-v3	2.9 GB	10 GB	1×	most accurate
turbo	1.5 GB	6 GB	~8×	the GPU default: large-v3 accuracy, fast; no translation

Speeds are relative to large-v3 on a GPU, from the Whisper paper.

Command line

With the Add the sorify command to PATH option, open a new terminal and:

sorify talk.mp3                                   text next to the file, auto model and language
sorify lecture.mp4 D:\Recordings -r -m turbo -l es -f txt srt
sorify interview.m4a --task translate -m medium   English translation
sorify podcast\ -o C:\Transcripts -f json --word-timestamps
sorify --help                                     every option
sorify --gui                                      open the window

Option	Meaning
-m, --model	auto, tiny … large-v3, turbo
-l, --language	en, es, Spanish, … (default: detect)
--task translate	translate speech into English text
-f, --format	one or more of txt timestamps srt vtt tsv json
-o, --output-dir	write everything here instead of next to each file
-r, --recursive	include sub-folders of the given folders
--overwrite	redo files that already have outputs
--device cuda|cpu	force the GPU or the CPU
--fast	greedy decoding instead of beam search
--word-timestamps, --max-line-width N, --prompt "…", --no-context	the Advanced options

Exit code 0 when every file succeeded, 1 when some failed, 130 when stopped with Ctrl+C.

Updates

Sorify checks GitHub Releases a few seconds after it opens (switch this off in About → Check for updates when the app starts; About → Check for updates still works by hand). When a newer version exists you see the release notes and choose Update now, Remind me later or Skip this version. Update now downloads the installer in the background, verifies its SHA-256 against the value GitHub publishes for the release, closes the app, runs a silent update and reopens Sorify. Updates never re-download PyTorch or ffmpeg unless the release says they changed.

Privacy

Your audio, video and transcripts are processed on your computer and never leave it. Sorify connects to the internet only to download its components, speech models and updates, and to ask GitHub for the latest version (an anonymous request that carries the app's version number). There is no telemetry, no account and no analytics.

Where things live

What	Where
The app, its private Python runtime, PyTorch, ffmpeg	%LocalAppData%\Programs\Sorify
Settings	%LocalAppData%\Sorify\settings.json
Logs (useful when reporting a problem)	%LocalAppData%\Sorify\logs (About → Open log folder)
Speech models	%UserProfile%\.cache\whisper

Uninstalling (Settings → Apps, or the Start menu entry) removes the app folder and asks whether to delete the settings. The model cache is left for other Whisper tools; delete it by hand to free the space.

Troubleshooting

Setup said the components were not installed. The download was interrupted or the network blocked it. Open Sorify: it offers to download them again. The log is in %LocalAppData%\Sorify\logs\setup.log.

The device chip says "CPU mode" although I have an NVIDIA GPU. Setup picks the GPU build when the NVIDIA driver reports CUDA 12.6 or newer and the GPU is a GeForce 10-series (Pascal) or later. Update the driver from NVIDIA, then About → Reinstall components. GeForce 9-series and older cards are not supported by current PyTorch builds.

"Not enough GPU memory". Pick a smaller model (small needs 2 GB, turbo 6 GB) or switch Run on to CPU under Advanced.

Wrong language detected. Set the language in the sidebar; detection looks at the first 30 seconds only.

The transcript repeats a sentence over and over. Switch off Carry context between chunks under Advanced and transcribe again.

Translation keeps the original language. turbo cannot translate; use medium or large-v3.

My antivirus flags the installer. Unsigned installers are sometimes flagged by heuristics. Compare the SHA-256 with SHA256SUMS.txt in the release; if it matches, the file is the one published here. If your scanner still complains, please open an issue with its name and the detection it reports.

Something else. Open an issue and attach sorify.log from the log folder.

Support the project

Sorify is free and made in my spare time. If it saves you hours of typing, a coffee keeps the updates coming:

Stars, bug reports and feature ideas are just as welcome.

License

Sorify is free to use, including commercially, under its End User License Agreement. It may not be redistributed, modified or reverse-engineered. This repository publishes the installer and documentation; the source code is not public.

Sorify downloads and uses open-source components under their own licenses: OpenAI Whisper (MIT), PyTorch (BSD), FFmpeg (LGPL), Python (PSF) and others listed in THIRD_PARTY_NOTICES.md.

---

Made by Anthony Fararoni · Changelog · All releases