The short answer
Live translation has two steps: speech recognition, then translation. Every tool here can run the first on your machine. Translation differs. Chrome sends the caption text to Google. Buzz translates into English on the computer and reaches other languages through an API endpoint, which can be a cloud service or a server you run. Sokuji and the OBS plugin LocalVocal translate locally or through cloud providers, and you choose. Voxxwire, Windows 11 on a Copilot+ PC, and Fraze, which is our app, keep both steps on the computer.
Then check what each one can hear. Chrome captions what plays in Chrome. Buzz hears a microphone unless you add a loopback driver. LocalVocal hears what OBS captures. Voxxwire, Sokuji's desktop app, Windows 11 and Fraze hear the whole machine.
The options side by side
| Platforms | Hears | Speech local | Translation local | Price | |
|---|---|---|---|---|---|
| Voxxwire | Windows 10 and 11, 64-bit | Mic and system audio | Yes | Yes | Free, labelled beta |
| Sokuji | Windows, macOS, Linux, Chrome, Edge | Mic and system audio. Extension: meeting tab | Yes in Local Inference mode | Yes in Local Inference mode, else cloud | Free, AGPL-3.0. Cloud use costs extra |
| Buzz | Windows, Apple silicon Macs, Linux | Files, links, mic, system audio via a loopback driver | Yes | To English yes. Others via an API endpoint, local or cloud | Free, MIT |
| OBS LocalVocal | Windows, macOS, Linux, inside OBS | Any OBS audio source | Yes | Yes, or cloud providers if you pick one | Free, GPL-2.0 |
| Chrome Live Caption | Windows, Mac, Linux, ChromeOS | Audio in Chrome only | Yes, Google says | No, caption text goes to Google | Free |
| Windows 11 Live Captions | Windows 11 | Any audio on the PC, or the mic | Yes, Microsoft says | Yes, Copilot+ PCs only | Included with Windows |
| Fraze | macOS 14 or later, Windows 11 | Any audio, one app on a Mac, or the mic | Yes | Yes | 30 minutes free, then paid |
All of it is from each project's own pages at the time of checking. Prices and hardware rules change, so look again.
The four apps, one by one
Voxxwire is a free app for Windows 10 and 11, and its own site marks it as a beta. It listens to your microphone and your system audio at once, subtitles both sides, and speaks the translation aloud. A local Whisper model does the recognition, Argos Translate does the translation, and language packs download on first use. The site lists 49 or more languages and calls the app open source. Its GitHub repository had no license file at the time of checking.
Sokuji is free and open source under AGPL-3.0, a two-way meeting translator that comes as a desktop app and a browser extension for Chrome and Edge. Its Local Inference mode runs recognition, translation and speech on your device, and the README describes it as fully offline with no API key. Sokuji also connects to cloud providers such as OpenAI, Google Gemini and Soniox, with your own key or a metered Kizuna AI account, so know which mode you are in. The README credits the two-way mode that detects the spoken language to Soniox, a cloud provider. The desktop app reads system audio and shows up as a microphone, so the call hears your translated voice.
Buzz is open source under the MIT license, from Chidi Williams. It transcribes files and YouTube links and records live from a microphone, and the docs show how to feed it computer audio through a loopback driver such as BlackHole or VB-CABLE. Its built-in translation is Whisper's own task, which only produces English. Other languages need an OpenAI-compatible endpoint. The docs say that can be a hosted model, at around a dollar for an hour of audio, or a local server such as Ollama or LM Studio.
LocalVocal is a free plugin for OBS Studio under GPL-2.0. It adds an audio filter that you put on a speech source, so it captions what OBS captures, and the project has a tutorial on captioning other applications. Recognition uses Whisper through whisper.cpp on your machine, and translation uses CTranslate2. The README says captions can be translated with local translation models, with Whisper's built-in translation, or with cloud providers. Its tutorials name DeepL and OpenAI.
What Chrome and Windows already give you
Chrome has Live Caption built in. Google's help page says audio and captions are processed locally and never leave your device. Translation is the exception: the same page says captions are sent to Google to be translated. Google describes the feature as captioning media from Chrome browser, so a desktop Zoom client or a game outside the browser is out of reach.
Windows 11 captions any audio passing through the PC. Microsoft's page says captions are generated on-device and that audio, voice data and captions never leave it. Translation is narrower: Microsoft says it needs a Copilot+ PC running Windows 11 version 24H2 or later, and that it covers over 40 languages into English and 27 into Chinese (Simplified).
Where Fraze fits
Fraze is our app, for macOS 14 or later on Apple silicon and for Windows 11. Unlike most of this list it is a paid, closed source product. It listens to everything the computer plays, to one app on a Mac, or to the microphone. Both steps run on the computer, so after the first language download it works with Wi-Fi off. Lines land about half a second to a second after a speaker pauses. A caption window can float over a call, and transcripts stay on the machine. Speech covers 25 European languages, and more on macOS 26 or later including Japanese, Korean and Chinese. Translation covers about 50 languages to or from English.
The first 30 minutes are free on every account, with no card. After that it is $29.99 once or $2.99 a month, both unlimited, with no per-minute meter. A free Fraze account carries the plan and never sees audio or transcripts. Fraze does not handle files, summaries, speaker labels, spoken translations or mobile. Verli, the cloud sibling from the same company, has summaries and sharing and runs in the browser, but its audio goes to the cloud.
Which one to pick
- Streaming or recording in OBS: LocalVocal.
- The other side has to hear you, not read you: Sokuji, which shows up as a microphone in the call app. Voxxwire also speaks translations aloud, but its pages do not say how that audio reaches a call.
- Recordings, speaker labels and subtitle files: Buzz. Fraze does none of those.
- Everything sits in a Chrome tab and Google may see the text: Chrome Live Caption.
- A Copilot+ PC, and you read English or Simplified Chinese: Windows 11, at no cost.
- Linux: Sokuji, Buzz or LocalVocal. Fraze does not run on Linux, Windows 10 or Intel Macs.
- Any audio on a Mac or a Windows 11 PC, both steps local: Fraze.
Questions
Which offline real-time speech translation app for desktop should I pick?
No single one, because they do different jobs. LocalVocal for streaming in OBS, Sokuji or Voxxwire for a spoken translation, Buzz for recordings, Fraze for reading any audio a Mac or a Windows 11 PC plays.
Is there an app that translates any audio playing on my computer without the cloud?
Yes, several. Voxxwire on Windows and Sokuji's desktop app capture system audio and can run everything locally. Windows 11 does too, but only on a Copilot+ PC and into English or Simplified Chinese. LocalVocal can too, inside OBS. Fraze captures system audio on macOS and Windows 11. Chrome does not fit: its translation goes through Google.
Which apps run speech models locally for live captions on Mac or Windows?
On both: Sokuji, Buzz, LocalVocal and Fraze. On Windows only: Voxxwire, and the Live Captions feature in Windows 11. Buzz needs an Apple silicon Mac, and so does Fraze, on macOS 14 or later. Chrome also generates its captions on the device, Google says.
Can I get unlimited live subtitles for foreign streams without paying per minute?
Yes. Run locally, none of these bills by the minute. Voxxwire, Sokuji, Buzz and LocalVocal are free downloads, and Chrome and Windows 11 cost nothing extra. Only the optional cloud providers in Sokuji, Buzz and LocalVocal can cost money. Fraze gives every account 30 minutes free with no card, then $29.99 once or $2.99 a month, both unlimited.
Does the translation step really run on my computer?
Check it per tool. Voxxwire: yes. Sokuji: yes in Local Inference mode, otherwise through the cloud provider you pick. LocalVocal: yes, unless you set a cloud provider. Buzz: yes into English, and other languages go through an endpoint that can be a server you run. Chrome: no, the text goes to Google. Windows 11: yes, on Copilot+ PCs only. Fraze: yes.