MacDub

Real-time, fully offline dubbing for macOS.

MacDub captures the audio of any app (Zoom, Teams, VLC, Safari, Chrome, Music — or the whole system), transcribes it, translates it and reads the translation aloud with a system voice, keeping the original quietly in the background like a documentary. Everything runs on your Mac with Apple's own frameworks. No cloud, no API keys, no virtual audio drivers, and the microphone is never touched.

Screenshots

The screenshots below were taken while MacDub dubbed Apple's WWDC25 session Bring advanced speech-to-text to your app with SpeechAnalyzer playing in Google Chrome, English → Spanish.


Dubbing — pick the app, the languages and a voice, press Start	Live transcript — original + translation, the word being spoken highlighted, latency per sentence
Floating subtitle bar — always on top of the video you are watching	Menu bar panel — status, last lines, Start/Stop and quick controls
Live translation (beta) — call app, microphone, both languages, a voice for each side, audio and/or chat	Live conversation — you and them side by side, a "…" bubble while someone speaks, transcription · audio · total per sentence
History › Dubbing — recorded audio with a live spectrum, the transcript following karaoke-style, playback modes, transcription · audio · total times per sentence, right-click to copy a line	History › Live translation — each conversation as a chat, both sides' audio in one take, times per sentence, right-click to copy a line
Subtitles — transcript options, floating bar and exports	AI & MCP — local summaries and one-click MCP registration
Settings — General, Capture, Speech, Voice, Subtitles, AI & MCP, Extensions, Permissions	Settings › Voice — quality badges and your own Personal Voice
Settings › Extensions — the Google Meet extension for live translation, one click to install

macOS 15 Sequoia or later (developed on 15.7.3, also built and run on macOS 26.7 / Apple M1).
Universal binary — Intel and Apple Silicon.
Builds without Xcode: the Command Line Tools are enough.
Interface in English, Spanish and Portuguese (Brazil); easy to add more.
MIT licensed. Repository: https://github.com/lordbasex/macdub
Created by Federico Pereira lord.basex@gmail.com

Features

Area	Feature	Details
Capture	One app or the entire system	Per-app capture follows helper processes (Chrome/Electron helpers, WebKit for Safari); "🔊 Entire system" captures everything except MacDub's own voice. Zoom, Teams, VLC, browsers, Music… The microphone is never captured.
	Original audio in the background	Core Audio process tap engine: the original is removed from the output and re-played at the level you choose (default 25 %), or lowered only while the translated voice speaks — documentary style.
	ScreenCaptureKit fallback	Used automatically when the app has no audio process yet (no background mix in that mode).
	Audio buffer for snippets	The last 60 s (configurable in Settings › AI & MCP, 0 = off) stay in memory so an assistant can request a WAV excerpt through MCP; never written to disk unless asked.
Recognition	On-device speech recognition	`SFSpeechRecognizer` with on-device models (macOS 15+).
	`SpeechAnalyzer` on macOS 26	Chosen automatically when the OS offers it; live fallback to `SFSpeechRecognizer` if it fails; compiled out on older toolchains. MacDub translates and speaks its finalized results — whole, corrected, punctuated sentences (85–92 % of sentences in one piece in the benchmark, against 35 % segmenting its volatile results) — and shows the volatile text live. The engine can be picked in Settings › Capture › Engine, which also says which one was detected and which one is running.
	Smart sentence segmentation	Cuts on punctuation, clause marks, length, pending time and silence; re-anchors when the recognizer rewrites earlier words. Unit-tested.
Translation	On-device translation	Apple's `Translation` framework, 21 languages, models downloaded once.
Live translation (beta, macOS 26+)	Talk through Meet, Zoom or any call app	What you say reaches the call translated, in the voice you choose (your Personal Voice included), through a virtual microphone (BlackHole, installed from MacDub with one click); what they say reaches your headphones in your language. Two SpeechAnalyzers at once with fast results: about 1.8 s from the end of a sentence to its translated voice.
	Google Meet extension	Switches Meet's microphone to the translated voice by itself (and back when you stop), posts your translation in the chat and translates the chat messages — Settings › Extensions.
	Conversation view	Chat-style, with a "…" bubble while someone speaks and, for each sentence, transcription · audio · total times; choose the microphone, the languages (those on your Mac first) and a voice for each side.
Voice	System voices with quality badges	🟢 Premium · 🟡 Enhanced · ⚪ Compact · ⚫ novelty; per-voice rate and volume; one-click reload after downloading voices. (Siri voices are not available to third-party apps.)
	Personal Voice	Dub with your own voice: the Personal Voice recorded in System Settings › Accessibility › Personal Voice appears first in the list (👤) once MacDub is allowed to use it (Settings › Voice › Personal Voice › Allow…). Apple supports it from macOS 14 Sonoma, so every macOS MacDub runs on (15+); creating the voice needs a Mac with Apple silicon, and apps cannot create one — Create in System Settings… opens the pane. It speaks the language it was recorded in.
	Latency management	Gentle speed-up when behind (configurable or off), stale sentences skipped past Max delay, backlog dropped with one click. Measured: translation ≈ 0.4 s, sentence end → voice ≈ 0.8 s with an empty queue, ≈ 2 s mean.
	Karaoke highlighting	The word being spoken is highlighted in the panel, the floating bar and the menu bar panel.
Subtitles	Transcript	While dubbing, the Dubbing screen becomes the live transcript: original + translation, per-sentence latency, level meter, skip/clear. The Subtitles screen holds the transcript options and exports; subtitles-only mode = voice off.
	Floating subtitle bar (⌘B)	Always-on-top pill: 1–6 lines, adjustable width and text size, drag handle, Stop/Start, quick settings, hide.
Menu bar	Status icon	Green dot while dubbing, amber while starting; click for a panel with the last lines, Start/Stop and quick controls; right-click for a menu.
	Menu-bar-only mode	Hide the Dock icon; open at login.
	Automation	Auto-start when the selected app starts playing audio; silence watchdog (notice or automatic stop).
Shortcuts	Global	⌃⌥D start/stop · ⌃⌥S subtitle bar (no Accessibility permission needed).
	In-app	⌘R start/stop · ⌘B subtitle bar · ⌘K skip backlog · ⌘L clear (starts a new session) · ⌘E export .srt · ⇧⌘E export .md · ⌘Y history · ⌘, settings.
Transcripts	Export	`.srt` (translation / original / both), `.md` with a session header (made for LLMs), `.txt`. File names carry no spaces (`macdub-transcript-<date>.<ext>`).
	Session history (⌘Y)	Every session archived when it stops, with the original audio recorded as AAC (`~/.macdub/audio`). Play it back with a live spectrum while the transcript follows karaoke-style — original audio only, documentary style (original lowered under the translated voice) or translated voice only, with original / translated / both texts, remembered between sessions; export audio + `.srt` under one name (`macdub-audio-<date>.m4a/.srt`) so VLC picks the subtitles up; delete a session and its audio goes with it. Settings › General shows the size of `~/.macdub` and offers Delete all sessions and Factory reset.
AI	Summaries	Claude Code and Codex CLIs, Ollama and LM Studio (choose the model), Apple Intelligence on macOS 26, or paste into Claude Desktop / ChatGPT — in the dubbing target language. A live seconds counter runs while it works; the summary shows how long it took and the tokens used (input/output; exact for Apple Intelligence on 26.4+, Claude Code, Ollama and LM Studio, ≈ estimated for Codex) plus the cost for Claude Code.
	MCP server	18 tools (including History: list/delete sessions, export audio + `.srt`, storage usage), live resources with push notifications, saved sessions as resources, 3 prompts; stdio and Streamable HTTP transports; server-side summaries with local models; one-click registration for Claude Code, Claude Desktop and Codex that survives moving the app.
Platform	Requirements	macOS 15 Sequoia or later (tested on 15.7.3 and 26.7), Intel and Apple Silicon (universal binary).
	System status	Settings › General › Status: chip, cores (performance + efficiency), memory, macOS version, the speech engine in use and Apple Intelligence availability (with the reason when it is off).
	Build	No Xcode needed — Command Line Tools with Swift 6; `make setup` installs them if missing. No third-party dependencies.
	Languages	Interface in English, Spanish and Portuguese (Brazil); adding one is copying a folder.
	Quality	Unit tests (Swift Testing), CI on GitHub Actions, release script with notarization and Homebrew cask.
	License	MIT.

Install

Homebrew

brew tap lordbasex/macdub https://github.com/lordbasex/macdub
brew install --cask macdub

Download

Grab MacDub-<version>.dmg from the Releases page, open it and drag MacDub to the Applications folder shown next to it (a .zip with the bare app is there too). Until releases are notarized, right-click › Open the first time.

Build from source (no Xcode needed)

git clone https://github.com/lordbasex/macdub.git
cd macdub
make setup      # checks macOS 15+, Command Line Tools with Swift 6, SDK frameworks; creates the dev signing cert
make check      # quick type-check (debug build, no bundle)
make run        # native-arch build → build/MacDub.app, then launches it
make            # universal (x86_64 + arm64) build
make zip        # universal zip to try on another Mac
make dmg        # universal build packed as a drag-to-Applications disk image (build/MacDub-<version>.dmg)
make test       # unit tests

make dmg needs only macOS tools (hdiutil, SetFile, osascript); the mounted image shows the MacDub icon, a background with an arrow, the app and an Applications shortcut. The Finder layout step asks once for permission to control Finder. make dmg-native packs the current build/MacDub.app without rebuilding.

There are no third-party dependencies — nothing is downloaded. The only requirement is macOS 15+ with the Command Line Tools (Swift ≥ 6.0). On a fresh Mac, make setup (or the first make run) launches xcode-select --install for you, waits for Apple's installer to finish, verifies the SDK and creates the dev signing certificate.

SDKs. The same sources build with the macOS 15, 26 and 27 SDKs, with Xcode or with the Command Line Tools alone. CI builds, tests and packages with all three on every push. The SDK decides which macOS 26 features are compiled in; at run time each of them is still checked (#available), so a build made with a newer SDK runs on macOS 15 too.

SDK	Toolchain (CI uses the Xcode listed)	SpeechAnalyzer	Apple Intelligence summaries	Notes
macOS 15.x	Xcode 16.4 · CLT 16 (Swift 6.0/6.1)	—	—	`SFSpeechRecognizer` only; everything else identical.
macOS 26.x	Xcode 26 · CLT 26 (Swift 6.2+)	✓ on macOS 26	✓ on macOS 26 (exact token counts on 26.4+)	Recommended.
macOS 27.x	Xcode 27	✓ on macOS 26+	✓ on macOS 26+	With the Command Line Tools only, the 27 SDK's `@State` macro needs a `SwiftUIMacros` plugin the CLT don't ship: `scripts/swift-flags.sh` then builds against the newest installed SDK that doesn't need it (e.g. 26.5) and prints which one.

Check which SDK a build uses with xcrun --show-sdk-version; force one with SDKROOT=$(xcrun --sdk macosx26.5 --show-sdk-path) make dmg. Settings › General › Status in the app says whether the running build has the macOS 26 features compiled in.

Signing for development. The build script signs with a self-signed "MacDub Dev" certificate if it exists (created by make setup / scripts/make-dev-cert.sh), otherwise ad-hoc. Ad-hoc identities change on every build and macOS then drops the Screen Recording grant, so the certificate is worth having.

First run

Screen & System Audio Recording — macOS asks on first launch. Enable MacDub in System Settings › Privacy & Security and relaunch the app (macOS only applies this grant after a relaunch).
Speech Recognition — requested when you first press Start.
Spoken language — pick one marked as on-device. If it says needs download, add the language under System Settings › Keyboard › Dictation so macOS fetches the model.
Translate to — the first time a language pair is used, Settings › Speech › Prepare translation downloads the model (needs internet that one time); MacDub also offers it when you press Start.
Voice — choose a voice and press ▶ to hear it. For better voices: Settings › Voice › Manage voices… → Accessibility › Spoken Content › System Voice › Manage Voices, download Enhanced/Premium voices, then ↻. To dub with your own voice, record a Personal Voice in the dubbing language and allow it in Settings › Voice › Personal Voice.
Pick what to capture (an app or 🔊 Entire system), play something and press Start (⌘R).

With the tap engine the app must already be producing audio when you press Start; otherwise MacDub tells you and falls back to ScreenCaptureKit (no background mix).

The app

One dark window with a sidebar of four screens, plus a standard Settings window (⌘,) and a menu bar item.

Screen	What is there
Dubbing	The session choices: application (or 🔊 Entire system), spoken language, target language, voice with ▶ test, original audio level, speak on/off, and the round Start button. While dubbing the hero is replaced by the live transcript (karaoke highlighting, latency per sentence, level meter, Skip / Clear).
Subtitles	Transcript options (include original, highlight spoken words, auto-scroll), the floating subtitle bar (⌘B) and the exports (.srt / .md / .txt).
History (⌘Y)	Saved sessions with a player: spectrum, transport, scrubber, and the transcript following the audio karaoke-style. The ⋯ icon on the player picks what you hear (original only · documentary: original low under the translated voice · translated voice only) and what you read (original · translation · both); the choice is remembered. Export audio + `.srt`, or transcripts; delete removes the audio too.
AI & MCP	Summaries with a local assistant (Claude Code / Codex CLIs, Ollama, LM Studio, Apple Intelligence) and one-click MCP registration for Claude Code, Claude Desktop and Codex.

Settings (⌘,) — General (interface language, menu-bar-only mode, open at login, global shortcuts, sessions & audio recording, storage size of ~/.macdub, delete all sessions, factory reset) · Capture (engine, original level, auto-start, silence watchdog) · Speech (recognition engine, silence cut-off, translation model) · Voice (voice, rate, volume, catch-up policy) · Subtitles (transcript and floating bar) · AI & MCP (HTTP transport, audio buffer for snippets) · Permissions.

Menu bar — green dot while dubbing; click for a panel with the last lines, Start/Stop and quick controls; right-click for a menu (start/stop, subtitle bar, settings, history, about, quit).

MCP server

Kind	Name	What it does
tool	`get_status`	phase, app, languages, engines, latency, silence, available apps
tool	`get_transcript`	current session as `md` (default), `txt`, `srt` or `json`; `content` original/translated/both; ranges: `last` N, `sinceIndex` (incremental reads — json returns `nextIndex`), `fromSeconds`/`toSeconds`
tool	`summarize_transcript`	summary produced by a local model on the Mac (Apple Intelligence on macOS 26, Ollama, LM Studio) so the transcript never enters the client's context; same range options, `sessionId` for saved sessions
tool	`list_local_models`	which local summarisation providers/models exist
tool	`search_transcript`	find sentences containing a phrase, with timestamps
tool	`list_sessions` / `get_session` / `delete_session`	saved sessions (with `audioPath` when the original audio was recorded) and their transcripts; delete removes transcript and audio
tool	`get_storage`	disk used by recorded audio (`~/.macdub`) and transcripts, sessions with/without audio — for clean-ups
tool	`export_session`	write the live or a saved session to a file (srt/md/txt, `path`, `overwrite`) and return its path; format `audio` copies a saved session's original audio as `<name>.m4a` next to `<name>.srt` so VLC loads the subtitles automatically (`macdub-audio-<date>` by default)
tool	`get_audio_snippet`	the last N seconds of captured audio as a 16 kHz mono WAV (file path, or embedded audio content with `inline`) — to double-check a sentence with an audio-capable model; MacDub keeps 60 s in memory by default (Settings › AI & MCP › Keep audio for snippets)
tool	`start_dubbing` / `stop_dubbing` / `clear_transcript`	control the app (`bundleIdentifier` or `system`; apps launched after MacDub are picked up, and the start is confirmed with the engines in use)
tool	`set_languages`, `set_voice`, `set_speak`, `set_original_volume`	change languages (while idle), voice, mute the dub, background level — live
resource	`macdub://transcript`, `macdub://status`	attachable live context; subscribable — the server pushes `notifications/resources/updated` as sentences arrive
resource	`macdub://sessions/<id>` (+ `/md	txt
prompt	`summarize`, `action_items`, `check_translation`	ready-made instructions over the live transcript

Register from the app (AI & MCP › Add to Claude Code / Claude Desktop / Codex) or by hand:

# stdio (recommended: Claude Code starts the server on demand). The launcher below is written by
# MacDub on launch and follows the app if you move it; the bundle path works too.
claude mcp add --scope user macdub "$HOME/Library/Application Support/MacDub/bin/macdub-mcp"
# HTTP (when the app serves it: Settings › AI & MCP › Also serve over HTTP)
claude mcp add --transport http --scope user macdub-http http://127.0.0.1:8765/mcp

{ "mcpServers": { "macdub": { "command": "/Users/<you>/Library/Application Support/MacDub/bin/macdub-mcp" } } }

How it works: the app writes ~/Library/Application Support/MacDub/live.json a few times per second and archives sessions in sessions/ (recorded audio goes to ~/.macdub/audio/<id>.m4a); the server reads those files and sends commands back through DistributedNotificationCenter. MacDub must be running for live data; saved sessions, storage and exports work even when it is not.

Transports. stdio by default. macdub-mcp --http 8765 [--token SECRET] serves the MCP Streamable HTTP transport at http://127.0.0.1:8765/mcp (POST for requests and batches, GET with Accept: text/event-stream for server notifications, optional bearer token, non-localhost Origin rejected). It binds to loopback only; for a remote assistant put it behind a tunnel or reverse proxy with TLS. The app can run it for you: Settings › AI & MCP › Also serve over HTTP on port.

Architecture

Sources/MacDubCore/            Pure logic, no UI frameworks, unit-tested:
                               Segment, TranscriptSegmenter, TranscriptExporter (srt/md/txt + cue timing),
                               Karaoke (word timing for playback), SpeechRate, LiveState,
                               SessionStore + MacDubPaths (~/.macdub)
Sources/MacDubMCP/main.swift   MCP server (JSON-RPC over stdio) → Contents/Helpers/macdub-mcp
Sources/MacDub/
  App/                         MacDubApp (scenes, menus), AppState (pipeline coordinator),
                               AppDelegate, StatusBarController (menu bar)
  Audio/                       ProcessTapCaptureManager (Core Audio tap, default),
                               AudioCaptureManager (ScreenCaptureKit fallback), AudioRingBuffer,
                               SessionAudioRecorder (AAC take per session), SessionPlayer (+ FFT spectrum)
  Speech/                      SpeechAndTranslationManager, RecognitionEngine (+ SFSpeechEngine),
                               SpeechAnalyzerEngine (macOS 26), TranslationBridge
  Voice/                       VoiceSynthesisManager (AVSpeechSynthesizer + backlog policy)
  UI/                          MainView (sidebar + sections), SettingsView (⌘,), Theme (building blocks),
                               SubtitlesView, FloatingSubtitlesView (pill), MenuBarPanelView,
                               HistoryView (player + karaoke transcript), AIIntegrationView, SpokenText
  Support/                     Settings, Localization, AIIntegration (summaries, MCP install),
                               AppleIntelligenceSummarizer, GlobalHotKey, LaunchAtLogin, errors
  Resources/<lang>.lproj/      Localizable.strings
Tests/Runner/                  Swift Testing suites + executable runner
Packaging/                     Info.plist template, entitlements, AppIcon.icns, dmg-background.tiff
graphics/                      logo.svg (vector twin of the app icon), pipeline.svg (the diagram above)
images/                        Screenshots and the rendered pipeline diagram used by this README
scripts/                       build-app.sh, run.sh, make-dev-cert.sh, make-icon.swift,
                               make-dmg.sh + make-dmg-background.swift, sign-and-notarize.sh,
                               release.sh, reset-permissions.sh
Casks/macdub.rb                Homebrew cask (updated by release.sh)
.github/workflows/ci.yml       Build, test, universal artifact, draft release on tags

Capture. Tap engine: CATapDescription(stereoMixdownOfProcesses:) over the app's audio processes (its pid, child processes, bundle-id prefixes, WebKit for Safari) or stereoGlobalTapButExcludeProcesses for the whole system, with muteBehavior = .mutedWhenTapped; the tap sits in a private aggregate device with the default output, and the IO callback copies input → output scaled by the chosen gain and produces a mono copy for recognition. ScreenCaptureKit engine: an audio-only SCStream with SCContentFilter(display:including:) / excludingApplications:, excludesCurrentProcessAudio = true.

Recognition & segmentation. A RecognitionEngine streams a continuously revised transcript. TranscriptSegmenter cuts it as soon as a sentence terminator appears, at a clause mark past 90 characters, at the last space before 160, after 4.5 s of pending text without a pause, or on silence; when the recognizer rewrites earlier words it aligns the words already emitted with the new text, so nothing is spoken twice. SFSpeechRecognizer gets its own rules: sentence ends also come from a pause in the audio followed by a stall in the text, a period at the end of a partial result waits for the next word, and a transcript the recognizer restarts inside a task is detected. Its runs are rotated on real silence, on errors, and at a pause 30–45 s in; a rotated run finishes its audio rather than being cancelled, so no words are dropped at the seam.

Translation. TranslationSession only exists inside SwiftUI's .translationTask; an invisible TranslationHostView keeps that closure alive as a job loop.

Voice. AVSpeechSynthesizer serializes utterances; MacDub adds a catch-up policy (rate boost per queued sentence, capped), stale-sentence skipping and backlog dropping, plus per-word highlighting through willSpeakRangeOfSpeechString.

History playback. The original audio of each session is recorded as mono AAC while dubbing (a session stopped and resumed keeps one file; the gap is padded with silence so the .srt cues stay aligned). SessionPlayer plays it through AVAudioEngine with an FFT spectrum; in the documentary and voice-only modes the translated voice is not a recording — the player speaks each translation with the same synthesizer as the live dub when its cue starts, so the highlighted word follows the real voice.

Benchmarks

MacDub's two speech engines were benchmarked on the same audio through the real dubbing pipeline — Apple M1, macOS 27, 2026-09-24 (30 minutes of speech, one engine at a time):

	SpeechAnalyzer (macOS 26)	SFSpeechRecognizer
Sentences spoken whole (one segment)	88 %	59 %
Word error rate	10.3 %	23.8 %
Words lost	1.4 %	6.8 %
Sentence ends found (recall)	90 %	44 %
Questions ending in `?`	18/25	10/22
Latency, median · p90 · max	0.7 s · 2.3 s · 9.3 s	1.0 s · 2.5 s · 8.9 s
Speech service CPU (avg)	5.7 %	27.6 %

SFSpeechRecognizer's segmentation was rewritten after 0.3.1. On the same audio it went from 14.7 % of the words lost and 55 % of sentences whole to 7.1 % and 61 %, with a worst latency of 8.9 s instead of 22.3 s (what changed).

SpeechAnalyzer now runs with fast results: its sentences come out three times sooner (median 2.3 → 0.7 s) with as many whole sentences (details). It waits for its finalized, corrected sentences; after 10 s without one it speaks what it has up to the last comma, trading a few whole sentences (92 → 85 %) for no long silences (details, macOS 26.7).

Run it on your Mac — especially M2, M3, M4, M5 and M6, which haven't been measured yet — and send the report as a pull request:

make benchmark                          # ~1 h: both engines, 30 min of audio each
scripts/benchmark/run.sh --minutes 10   # quicker

It writes docs/benchmarks/<date>-<chip>.md with the machine, the results and the method.

Tests

make test

Swift Testing suites over MacDubCore (segmentation rules, catch-up policy, exporters and cue timing, karaoke word timing, session records). They run as an executable because swift test needs Xcode's xctest loader, which the Command Line Tools don't ship; Package.swift adds the CLT Testing.framework paths only when Xcode is absent.

Translations

The UI follows the system language (or Settings › General › Language). To add a language: copy Sources/MacDub/Resources/en.lproj to <code>.lproj, translate the values in Localizable.strings (keys are the English text), add a case to InterfaceLanguage in Support/Localization.swift, run plutil -lint and make run. Dynamic strings go through L("…") / LF("…", args); SwiftUI literals localize on their own.

Releasing

VERSION=0.2.0 CODESIGN_IDENTITY="Developer ID Application: … (TEAMID)" NOTARY_PROFILE=MacDub-Notary make release

Builds the universal app, signs and notarizes (when the identity and notary profile are set), zips it, packs the .dmg, writes the SHA-256 of both, updates Casks/macdub.rb and creates the GitHub release with gh. The CI workflow builds and tests on every push and attaches an unsigned universal zip as a draft release on tags.

Known limitations

Latency of roughly 1–3 s is inherent to sentence-by-sentence dubbing.
On-device speech languages are limited to those with a downloaded Dictation model (usually the system language plus en-US).
SFSpeechRecognizer punctuates poorly and misses about 7 % of the words in the benchmark (the recognizer itself, not the pipeline); on macOS 26 SpeechAnalyzer is used instead when available.
With SpeechAnalyzer a sentence is spoken once it is complete — about 0.7 s after it ends (median), in whole sentences. When SpeechAnalyzer holds a sentence back (it sometimes merges two), MacDub speaks it after at most ~10–12 s.
Safari plays audio through shared WebKit XPC processes; tapping Safari can affect other WebKit apps.
No speaker diarization: everyone gets the same voice (the data model has a speaker field ready for it).
Siri voices cannot be used by third-party apps.
Live translation needs macOS 26 (two SpeechAnalyzers at once; two SFSpeechRecognizers cannot run in one app), the BlackHole virtual audio driver for the call to hear you, and headphones (with speakers, your microphone would send their translated voice back). Its chat features work with Google Meet in Chrome only, and the Meet extension installs in developer mode until it is in the Chrome Web Store.
History playback highlights the original text by spreading each sentence over its words (the recognizer keeps no per-word timing); the translated text follows the voice exactly.

Roadmap

Done

Later

Speaker diarization (voice per speaker) when Apple exposes it or a lightweight on-device model fits.
MacDub's own virtual microphone ("MacDub Mic"): a Core Audio driver (Audio Server Plug-in) in its own repository, so live translation no longer needs BlackHole installed from Homebrew and the call app shows MacDub's name.
Live translation chat beyond Meet in Chrome: Zoom, Teams and Slack through the macOS Accessibility APIs.
A third recognition engine running Whisper or Parakeet on Apple silicon with MLX: better accuracy and punctuation than SFSpeechRecognizer, the only engine on macOS 15. Fed in windows (not streaming), so it needs the latency measured. Proof of concept first: check that mlx-swift builds without Xcode (its Metal shaders), or else run MLX in a helper process, then compare it with the other two engines on the benchmark.
MacDub's own on-device neural voice, offered to the whole Mac as a speech synthesis provider (AVSpeechSynthesisProviderAudioUnit, an AUv3 extension, macOS 13+) with a local model such as Piper or Kokoro, run with MLX (mlx-audio has TTS models ready for Apple silicon). Start with a proof of concept: time to first audio on an M1 and quality in Spanish, and packaging the .appex without Xcode.
Keep summaries (with their time and tokens) in History next to each session.
Serve MCP HTTP requests concurrently (a long summary blocks other clients today).
More UI languages from the community.

Debugging

/usr/bin/log stream --predicate 'subsystem == "com.lordbasex.MacDub"' --level info

(log is also a zsh builtin — the full path matters.) Categories: app, capture, speech, translation, voice.

License

Author

Created by Federico Pereira — lord.basex@gmail.com · github.com/lordbasex

Navigation Menu

Repository files navigation

MacDub

Screenshots

Features

Install

Homebrew

Download

Build from source (no Xcode needed)

First run

The app

MCP server

Architecture

Benchmarks

Tests

Translations

Releasing

Known limitations

Roadmap

Done

Next

Later

Debugging

License

Author

About

Releases

Packages

Contributors

Languages

Name		Name	Last commit message	Last commit date
.github/workflows		.github/workflows
Casks		Casks
Packaging		Packaging
Sources		Sources
Tests/Runner		Tests/Runner
docs		docs
extensions/meet-chat		extensions/meet-chat
graphics		graphics
images		images
scripts		scripts
.gitignore		.gitignore
LICENSE		LICENSE
Makefile		Makefile
Package.swift		Package.swift
README.md		README.md

Latest commit

History

Folders and files

Repository files navigation

MacDub

Screenshots

Features

Install

Homebrew

Download

Build from source (no Xcode needed)

First run

The app

MCP server

Architecture

Benchmarks

Tests

Translations

Releasing

Known limitations

Roadmap

Done

Next

Later

Debugging

License

Author

About

Topics

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages