Real-time subtitles over everything, from your speakers or your microphone.
Omni-STT listens to sound - what your speakers are playing, what your microphone hears, or several of those at once - recognizes the speech and draws it as subtitles on top of whatever is on screen. The overlay is transparent and click-through, so it sits over a video, a game or a call without getting in the way of either.
Useful when the audio is bad, when the room is loud, when the speaker is hard to follow, or any time you would rather read than listen.
Grab a build from the Releases page and run it. There is no installer - the binary is self-contained, put it wherever you like.
Antivirus software sometimes flags it, because drawing a click-through window on top of everything is also what screen-capture malware does. Nothing can be done about that from this side.
- Open Settings → Speech Engine and pick a provider. Soniox needs an API key and works immediately; Vosk needs a download and works offline. Details below.
- Open Settings → Audio and tick what to listen to. Nothing ticked means your default output device, which is usually what you want.
- Close settings. Subtitles appear when someone starts talking.
The tray icon brings settings back and closes the app.
The list in Settings → Audio shows every device the system has, grouped into System audio - what a device is playing - and Microphone - what it hears. Tick as many as you like; they are mixed together and recognized as one stream.
Two useful combinations:
Your microphone plus system audio. In a call, the other people come through your speakers, and you go into your microphone, so ticking both transcribes the whole conversation rather than half of it. Wear headphones - otherwise your microphone picks up the speakers as well and every remote phrase is transcribed twice, a beat apart.
A virtual audio cable. Route one application (a voice chat, say) to a virtual cable such as VB-Cable or VoiceMeeter, and tick only that cable. The recognizer then hears voices and nothing else - no game audio, no music, no notification sounds. This is by far the best transcription quality you can get, and as a bonus the silences between phrases are real silences, so nothing is sent to a paid provider while nobody is speaking.
If you route audio to a cable you will stop hearing it yourself; send it back to your headphones with VoiceMeeter, or with Listen to this device on the cable's recording side. If you do the latter, leave your headphones unticked - they now carry the same audio as the cable, and ticking both transcribes everything twice.
On Windows and Linux, system audio capture works without extra software — on Linux, pick the PulseAudio monitor source, which shows up in the Microphone group. macOS has no built-in loopback at all and needs a virtual device such as BlackHole; that route is untested.
Soniox: cloud recognition, accurate, and the only one of the two that labels who is speaking and can translate as it goes. Needs an API key, which is kept in the operating system's keychain (Windows Credential Manager, macOS Keychain, Linux Secret Service) and never written to the config file. Audio leaves your machine.
Vosk: runs entirely on your computer. No account, no network, nothing leaves the machine. Less accurate, and no speaker labels. The engine and the language model are not bundled - together they are hundreds of megabytes, and most people only ever use one of the two providers.
Two downloads. Settings tells you whether it found each of them.
-
The engine library. Take the build for your platform from the Vosk releases and unpack it. Put
libvosknext to the Omni-STT executable, or point at it in Settings → Speech Engine → Library. On Windows keep thelibgcc,libstdc++andlibwinpthreadDLLs from the same archive beside itlibvosk.dllwill not load without them. -
A language model. Pick one from the model list and unpack it. Point at the unpacked folder - the one containing
am/andconf/- in Settings → Speech Engine → Model.
Models run from about 50 MB to well over a gigabyte. The small ones load in seconds and are usually enough for live subtitles.
Turn on Settings → General → Save transcripts and every finished line is
appended to transcripts/omni-YYYY-MM-DD.txt next to the executable, with the
time of day and the speaker where the provider gives one:
=== 2026-09-22 14:03 ===
14:03:17 [1] hey, can you hear me?
14:03:24 [2] yes, you're fine
14:03:31 [1] then let's start
A new file per day, in your local time. Plain text, so grep works.
Most of the settings say what they do. Three do not:
Threshold is how loud sound has to be before it counts as speech. Below it, audio is dropped and never sent anywhere, which saves money on a paid provider and keeps stale subtitles from lingering. Too high and quiet speech is ignored; too low and background noise keeps the stream permanently awake. A microphone usually needs a lower value than system audio.
Hangover Chunks is how long the stream stays open after the sound drops below the threshold, so a pause mid-sentence does not cut the phrase in half.
Max Blocks is how many blocks can the current overlay contain. One block is one speaker.
Max Lines is how many lines can a single block consist of.
Needs the Rust toolchain. Linux also
needs ALSA, GTK 3 and libxdo development packages, see the apt list in
.github/workflows/rust.yml for the exact names.
git clone https://github.com/eoftgge/omni_stt.git
cd omni_stt
cargo build --release
The binary lands in target/release/.
MIT, see LICENSE. Bundled fonts and dependencies have their own terms, listed in THIRD-PARTY-NOTICES.md.
Open an issue. Attaching
logs/omni.log (turn on Settings → App → Log to file first) saves a
round of questions.
If Omni-STT is useful to you and you'd like to support its development:
| Coin | Network | Address |
|---|---|---|
| USDC, USDT or SOL | Solana | 9yjnQRWAEifMnZHSNQKcr1Drwzw8H9on3cZXien4duQ |
| BTC | Bitcoin | bc1qzdq6pdxzwu9tyqepkgdx0k0tndape9ww3rynr2 |
Please double-check the network before sending: coins sent on the wrong network are lost. Copy the address from this page, not from your transaction history.
