Ben

Real-time English ⇄ German speech translation on macOS. Fully on-device — no API keys, no model downloads from this app, no network round-trips.

Pipeline: AVAudioEngine → SFSpeechRecognizer → Translation framework. All Apple system frameworks; everything runs locally.

---

Download

Pre-built DMGs are attached to each GitHub Release. Open the DMG, drag Ben.app to Applications, then right-click → Open the first time so macOS lets the ad-hoc-signed bundle launch.

Build from source

git clone git@github.com:alileza/ben.git
cd ben
./build.sh
open Ben.app

build.sh produces an ad-hoc-signed bundle for local dev. For a release that other people can open without right-click→Open gymnastics, see release.sh.

First launch will ask for Microphone and Speech Recognition permission, and (on first use of a language pair) prompt to download the on-device translation model.

Requirements

Requirement	Notes
macOS 15+ (Sequoia / Tahoe)	The Translation framework needs 15
Xcode + Swift 6 toolchain	swift --version should report 6.x
Dictation enabled	System Settings → Keyboard → Dictation. Without it SFSpeechRecognizer errors with "Siri and Dictation are disabled"
EN ⇄ DE translation pair	macOS prompts to download on first use

Features

• Two-column transcript with a shared scroll — source and translation rows are always vertically aligned. Active in-progress row is pinned at the top under the status bar.
• One-click direction toggle (EN ⇄ DE).
• Input device picker + hardware mic volume in a popover next to the start/stop button.
• Diagnostics pane: live mic peak, translation latency chart, 30 s sliding window.
• Debug log pane: in-app ring-buffered events, mirrored to os.log under subsystem com.local.ben so log stream works too.
• Transcript export to .txt — source only, translation only, or both paired with timestamps.
• Smart chunking: pauses commit a new line; long monologues chunk at word boundaries after 5 s (10 s hard cap).
• Canonical translation pairing: the committed row's source and translation always correspond — no race conditions where the displayed translation lags behind the recognizer's final text.

Project layout

Sources/Ben/
├── App.swift                @main entry + View-menu commands
├── AppState.swift           Observable shared state + translation queue
├── Models.swift             Direction, PairedRow, MicPoint, LatencyPoint
├── Audio/
│   ├── AudioEngine.swift    AVAudioEngine → AsyncStream
│   └── AudioDevices.swift   CoreAudio HAL enumeration + volume control
├── Speech/
│   └── SpeechEngine.swift   SFSpeechRecognizer with stale-callback guard
├── Export/
│   └── TranscriptExport.swift  .txt save via NSSavePanel
├── Logging/
│   └── DebugLog.swift       Ring buffer + os.log
└── Views/
    ├── ContentView.swift    Root composition + engine driver
    ├── StatusBar.swift      Top bar + pills + popover controls
    ├── Transcript.swift     Two-column transcript + active row
    ├── Diagnostics.swift    Mic + latency charts (Swift Charts)
    └── Debug.swift          Debug log pane

The architecture is documented in BEST_PRACTICES.md.

Watching debug logs

In-app: View → Show Debug Logs (⌥⌘D).

In a terminal:

log stream --predicate 'subsystem == "com.local.ben"' --style compact

Build & regenerate the icon

swift make-icon.swift   # writes AppIcon.icns + AppIcon.iconset/
./build.sh              # rebuilds the bundle with the new icon

If the Dock or Cmd-Tab shows a stale icon after a rebuild:

killall Dock Finder

Troubleshooting

Symptom	What to check
"Siri and Dictation are disabled" in the debug log	Enable Dictation in System Settings → Keyboard
Source pane fills but translation stays empty	Accept the macOS download prompt for the language pair on first use
App in /Applications doesn't appear in Spotlight or Alfred	Ad-hoc-signed apps in /Applications are filtered by Gatekeeper. Run sudo spctl --add /Applications/Ben.app or keep the build under your home dir
Cmd-Tab shows a stale white-square icon	killall Dock to bust the icon cache
German is recognized but onDevice=false in the log	Your Mac doesn't have on-device German; recognition falls back to the network

Limitations

• Speaker labels stay at S1; no real diarization.
• On-device speech availability varies by language and Mac model.
• Auto-detect (speak either language without flipping the toggle) is on the roadmap, not in the app yet.

Contributing

See CONTRIBUTING.md. PRs and issues welcome — keep them small and focused.

License

MIT. See LICENSE.