JaSub — Real-Time Meeting Transcription & Translation

Also available on iPhone & iPad: LiveSub →

Real-time speech transcription and translation for macOS 26.4+. No API keys, no subscription. Fully on-device by default; optional high-fidelity mode routes translation through Apple servers via Private Cloud Compute.

Uses Apple's native stack exclusively:

• ASR: SpeechAnalyzer + SpeechTranscriber (Speech.framework, macOS 26+)
• Translation: Translation.framework .highFidelity (Apple servers, optional) or .lowLatency (on-device, < 100ms, default) → FoundationModels / Apple Intelligence (fallback, ~500ms)
• System audio: CATapDescription + IOProc — captures browser or any app audio without a virtual audio driver

→ JaSub 0.3.9 (DMG)

Requires macOS 26.4+ (Tahoe) + Apple Intelligence + Apple Silicon.
First launch: right-click the app → Open to bypass Gatekeeper.

When replacing an existing installation, TCC (privacy permissions) must be reset manually — the ad-hoc signature changes with each build, and macOS does not transfer permissions automatically.

1. Open System Settings → Privacy & Security → Screen Recording → remove JaSub
2. Open System Settings → Privacy & Security → Microphone → remove JaSub
3. Drag the new version into Applications (overwrite)
4. Launch JaSub — macOS will prompt for permissions again → grant them
5. Quit and relaunch JaSub — TCC changes don't take effect in the same process session

A native SwiftUI menu bar app with floating subtitle overlay windows.

• macOS 26.4+ (Tahoe) with Apple Intelligence enabled
• Apple Silicon Mac
• Xcode 26+ (xcode-select --install or full Xcode)
• xcodegen (brew install xcodegen)

System Settings → General → Language & Region → Translation Languages

Install the language pairs you need (e.g. English ↔ Chinese, Japanese ↔ Chinese). Required for < 100ms translation. Without this the app uses FoundationModels (~500ms), which is still fully on-device.

cd app
./setup.sh        # install xcodegen and generate .xcodeproj (first time only)
./package.sh      # Release build → dist/JaSub-0.3.9.dmg

Open the DMG, drag JaSub to Applications, then right-click → Open on first launch (bypasses Gatekeeper — the app is ad-hoc signed, not notarised).

Both subtitle windows can be freely resized by dragging any edge or corner. Font size (12–40pt) is controlled from the settings popover and persisted across launches.

The UI adapts to the system language automatically. Supported languages:

English · 繁體中文 · 简体中文 · 日本語 · 한국어 · Français · Deutsch · Español · Português (BR) · Italiano · العربية · Русский · Nederlands · Polski · ภาษาไทย · Türkçe · Українська · Tiếng Việt · Bahasa Indonesia

Because JaSub is ad-hoc signed, macOS treats each new binary as a different app and revokes the old permissions. Follow these steps to avoid "Screen Recording denied" errors after upgrading:

1. Remove old permissions — System Settings → Privacy & Security:
   • Screen Recording → select JaSub → click – to remove
   • Microphone → select JaSub → click – to remove
2. Install the new version — drag the new JaSub.app into Applications (overwrite)
3. Launch JaSub — macOS will ask for Screen Recording and Microphone access again → grant both
4. Relaunch when prompted — macOS will say the app needs to be restarted for the permission to take effect → quit JaSub completely and reopen it
5. JaSub should now work normally

Why this is needed: TCC (macOS's permission database) records permissions by binary path + code signature. An ad-hoc re-signed binary is treated as a new app, so the old entry must be removed before macOS will prompt again.

• Speech Recognition — prompted on first start
• Microphone — prompted on first start (microphone mode)
• Screen & System Audio Recording — required for browser audio capture (System Settings → Privacy & Security → Screen Recording)

Microphone
    ↓ AVAudioEngine (16kHz mono float32)
    ↓ AsyncStream<[Float]>

Browser / System Audio
    ↓ CATapDescription + IOProc (48kHz → resample → 16kHz)
    ↓ AsyncStream<[Float]>

SpeechAnalyzer + SpeechTranscriber  [ASRManager]
    partial → live subtitle update
    final   → hallucination filter → translation queue

Translation.framework  [TranslatorManager]
    ① Translation.framework  .highFidelity  (Apple servers, optional)
       or .lowLatency  (on-device, < 100ms, default)
       .highFidelity failure → auto-fallback to .lowLatency
    ② FoundationModels fallback  (Apple Intelligence 3B, ~500ms)
       guardrail → completion-style retry → ⚠ original text

@MainActor TranslationEngine → SwiftUI floating NSPanel windows

SpeechAnalyzer (macOS 26+) is Apple's API for continuous live transcription. No session time limit, progressive partial results, fully Neural Engine, automatic model download.

Evolution: mlx-whisper → WhisperKit → SFSpeechRecognizer → SpeechAnalyzer. WhisperKit with --use-prefill-prompt invalidated the Neural Engine cache (7–10s/segment). SFSpeechRecognizer was superseded when SpeechAnalyzer shipped in macOS 26.

Music / singing: SpeechAnalyzer is speech-optimised. Some lyrics are recognised but gaps are expected.

Proper nouns: No hot-words API — character names and technical terms may be phonetically substituted.

A Python + Swift CLI version is available in server/ for headless or terminal use. No further development planned — the Menu Bar App is the primary version.

./run.sh    # auto-setup + interactive menu (language pair + audio source)

Requires Python 3.13+ and Xcode Command Line Tools. See NOTES.md for architecture details and known issues.