TalkThrough

Live captions and translation for any audio playing on your Mac. Captures system audio with ScreenCaptureKit, transcribes it on-device using macOS 26's SpeechTranscriber, and translates each line via DeepL — all rendered in a floating overlay window that stays on top of every Space.

Useful for watching foreign-language video or sitting in a meeting whose audio you don't speak.

Requirements

• macOS 26.0 or newer (uses SpeechTranscriber, introduced in the macOS 26 Speech framework).
• Xcode 26 or newer to build.
• A free DeepL API key — sign up at https://www.deepl.com/pro-api.
• An Apple Developer account (free tier is fine) for code signing during development.

Setup

git clone https://github.com/greg7gkb/TalkThrough.git
cd TalkThrough
cp Config/Secrets.example.xcconfig Config/Secrets.xcconfig
# Edit Config/Secrets.xcconfig and paste your DeepL API key
open TalkThrough.xcodeproj

Build and run from Xcode. The first launch will prompt for:

• Speech Recognition — required to use SpeechTranscriber.
• Screen & System Audio Recording — required for ScreenCaptureKit to capture audio playing through other apps.

The first transcription session also downloads the speech model for your locale (a few seconds to a minute, one-time).

Configuration

The target translation language is hard-coded to Spanish ("ES") in ContentView.swift for now. Change it to any DeepL target language code — "EN", "FR", "DE", "JA", etc. — to translate into something else. A picker is on the roadmap; see PLAN.md.

How it works

ScreenCaptureKit ──audio buffers──▶ SpeechTranscriber ──text──▶ ViewModel
                                    (macOS 26 Speech                │
                                     framework)                     ▼
                                                                 DeepL
                                                                    │
                                                                    ▼
                                                            floating overlay

• AudioCaptureEngine opens an SCStream on the main display, captures audio only (video is throttled to 1fps at 2×2 to keep ScreenCaptureKit happy), and forwards AVAudioPCMBuffers.
• SpeechEngine wraps SpeechAnalyzer + SpeechTranscriber with the progressiveTranscription preset. It resamples the audio to whatever format the transcriber wants, runs the results stream, and emits a running transcript by tracking results keyed by audio time range.
• LiveTranslateViewModel drives both engines and debounces translation calls: partials wait 500ms in case more text arrives, but a max-wait clamp forces a fire after 1s so continuous source audio doesn't starve the translator.
• TranslationService is a thin DeepL REST client using header-based auth (Authorization: DeepL-Auth-Key).

See PLAN.md for the prior phases, current state, and what's next.

Privacy and data

• Audio capture is local; raw audio buffers are sent to Apple's SpeechTranscriber (which transcribes on-device on macOS 26+) and then discarded.
• Transcribed text is sent to DeepL over HTTPS for translation. No other network calls.
• Your API key lives only in Config/Secrets.xcconfig, which is gitignored.

License

MIT — see LICENSE.