MacParakeet

Fast, private, local-first voice app for Apple Silicon Macs. Free and open-source.

There are many voice transcription/dictation apps, but this one is mine.

macparakeet.com

---

MacParakeet combines system-wide dictation, file/media transcription, and meeting recording in one local-first app, with optional selected-text Transforms and agent automation through macparakeet-cli. Parakeet v3 is the standard-path local speech engine; locale-aware first-run setup selects WhisperKit instead when the Mac has no preferred English language and prefers Korean, Japanese, Chinese, or Cantonese. Parakeet v2/Unified, Nemotron Beta, Cohere Transcribe, and WhisperKit remain selectable local choices for different language, latency, timestamp, and accuracy needs. Version 0.8.7 is the current stable DMG: hold-to-talk works again, Fn works with Caps Lock on, skip the microphone during first-run if you only transcribe files, keep AI Formatter off unless you turn it on (separate dictation and transcript prompts), optionally type finished dictation at the caret, and connect Moonshot, DeepSeek, Qwen, Z.AI, or MiniMax. Bundled CLI is 4.4.0. All speech recognition happens on your Mac; networked AI features are separate and opt-in.

Release status

The notarized DMG is the stable release channel.

See the canonical release and feature-gate status. App-aware AI Formatter profiles, activity-based meeting detection, and the in-process MLX LLM remain gated; source presence does not mean they are available in the stable app.

Meeting calendar support is live in the stable DMG. MacParakeet reads upcoming meetings from the local macOS Calendar store through EventKit, including Microsoft 365 and Exchange calendars enabled in System Settings → Internet Accounts. It can show reminders and optionally start a recording after a countdown. Auto-start defaults to .off and must be opted into. Recordings stop manually unless the separate activity-based auto-stop setting is enabled; that setting also defaults off.

What it does

The feature descriptions below cover current source. Gated or unreleased surfaces are called out in the canonical status table.

Dictation — Press a hotkey in any app, speak, text gets pasted. Hold for push-to-talk, or tap the hands-free shortcut to start and stop longer dictations. Works system-wide. The default uses Fn, but Settings -> Dictation can bind external-keyboard-friendly keys such as F13/F19 or End, modifier+key shortcuts, or modifier-only chords like Control+Option. A beta setting can pause supported Now Playing media while you dictate and resume it when capture stops.

File & URL transcription — Drag one or many audio/video files, drop a folder, use the multi-select picker, or paste a supported media or podcast link. YouTube, X, Vimeo, TikTok, Instagram, Facebook, and other sites handled by yt-dlp are supported subject to upstream site changes; the card recognizes known platforms as you paste. Apple Podcasts links resolve through the iTunes lookup API to the episode's audio enclosure (no scraping), then download and transcribe locally just like a YouTube video. The CLI also does freetext podcast search — macparakeet-cli transcribe --podcast "Lex Fridman episode 400" searches the iTunes directory, parses the show's RSS feed, picks the episode, and transcribes it. Local-file batches run sequentially, keep finished results in the Library, and can be cancelled as a group. Full transcript output includes word-level timestamps and speaker labels when the selected speech engine provides timings; Cohere produces plain text only. Completion chime/banner and export to 8 formats (TXT, Markdown, SRT, VTT, DAPT, DOCX, PDF, JSON) are supported. Assign global hotkeys to trigger File or URL transcription from anywhere.

Meetings — Record system audio and microphone together, or pick microphone-only or system-only capture (microphone-only needs no Screen Recording permission). See a live local transcript preview, take notes during the call, then save the finalized transcript to the library with export, prompts, and chat. Import an existing audio or video recording as a normal meeting while leaving the original file unchanged, or split a saved meeting into independently owned parts. Choose how long MacParakeet keeps its managed meeting audio: keep it, auto-delete after a set number of days, or remove it right after transcription.

Meeting calendar support — Grant Calendar access to get local reminders for upcoming meetings or opt into auto-start. MacParakeet uses calendars already configured on this Mac through EventKit; Microsoft 365 and Exchange work when Calendar is enabled for the account in System Settings → Internet Accounts. Settings can open that pane and refresh the visible calendar list. Skip auto-record for one occurrence or a repeating series without turning auto-start off globally. MacParakeet does not add Google or Microsoft sign-ins and cannot read calendars stored only inside Outlook. Recordings stop manually unless you separately enable activity-based auto-stop.

Text cleanup — Filler word removal, custom word replacements, text snippets with triggers. Deterministic pipeline, no LLM needed.

AI features — Optional summaries, chat, AI formatter, and Transforms for rewriting selected text through your configured provider. Connect a cloud provider (OpenAI, Anthropic, Gemini, OpenRouter, Moonshot/Kimi, DeepSeek, Qwen, Z.AI, MiniMax), local runtime (Ollama, LM Studio), Apple Intelligence on macOS 26 when the Mac is eligible, an OpenAI-compatible endpoint, or a CLI tool (Claude Code, Codex). Entirely opt-in.

Limitations

• Apple Silicon only (M1 or newer)
• Parakeet is best for English and its 25 supported European languages (v3 auto-detects among them); the v3 default is not a CJK/Korean engine
• Nemotron is Beta while real-world quality is benchmarked
• Nemotron, WhisperKit, and Cohere Transcribe require separate local model downloads before first use
• Cohere is batch-only: it can be used for dictation after recording stops, file transcription, and meeting finalization, but it does not show live dictation preview or meeting live-preview chunks and does not provide word timestamps/speaker labels

ASR benchmarks

The current ASR benchmark lives in benchmarks/asr/. It scores every engine through the same normalizer, uses full LibriSpeech test-clean + test-other for English, uses capped FLEURS slices for multilingual coverage, and reports hardware-specific speed/memory on an Apple M4 Pro with 48 GB RAM on macOS 15. Run benchmarks/asr/run_all.sh verify to re-score the committed hypotheses and validate the benchmark contract.

English accuracy, full LibriSpeech sets:

Multilingual coverage, FLEURS first 150 utterances per language. English is WER; Korean, Japanese, and Chinese are CER:

Speed and memory, same Apple M4 Pro micro-benchmark:

Cohere is the most accurate on-device engine in this benchmark, but its statistically clear wins are noisy English and Japanese. Clean English, Korean, and Chinese are ties with the best alternative under the paired-bootstrap test. Parakeet remains the default because it is fast, low-memory, timestamped, and strong on supported languages. Cohere is opt-in for accuracy-critical batch work on 16 GB+ Macs. The Cohere speed/memory row is a reference measurement; see benchmarks/asr/ for method notes.

Get it

Download: Grab the notarized DMG or visit macparakeet.com. Drag to Applications, done.

On the standard path, first launch downloads the default Parakeet CoreML build (~465 MB) plus speaker-detection assets (~130 MB) as needed. Locale-aware Korean/Japanese/Chinese/Cantonese setup downloads WhisperKit instead when no preferred English language is present. Parakeet v2 and v3 cache independently if you install both. Core dictation, local-file transcription, and meeting recording can work offline after required models are installed; media imports, updates, telemetry, and cloud/remote AI providers still require a network.

The DMG is the stable release.

Mac app (Homebrew cask):

brew install --cask macparakeet

This is the official homebrew/cask entry — no tap required. It installs the same notarized DMG as the direct download, and in-app updates continue through Sparkle.

Standalone CLI (Homebrew):

brew install moona3k/tap/macparakeet-cli
macparakeet-cli --version
macparakeet-cli health --json

The Homebrew formula installs the public macparakeet-cli surface plus Homebrew-managed ffmpeg and yt-dlp. It shares the same local database and model cache as the app.

Build from source:

git clone https://github.com/moona3k/macparakeet.git
cd macparakeet
swift test
scripts/dev/run_app.sh    # build, sign, launch

The dev script creates a signed .app bundle so macOS grants mic and accessibility permissions. It disables target-level Xcode signing, then signs the finished bundle with the best available local identity. Override with MACPARAKEET_CODESIGN_IDENTITY="Your Identity" if needed.

Command line and agent automation

macparakeet-cli is the public automation surface for MacParakeet: the canonical Swift-native interface to Parakeet TDT on Apple Silicon, plus the scriptable entry point for MacParakeet's local library, model cache, prompts, meetings, and JSON contracts. Use integrations/README.md for the agent-facing automation guide and Sources/CLI/CHANGELOG.md for compatibility notes.

Discover the installed binary's machine-readable command catalog. The examples describe current development; the stable bundled CLI and independently released Homebrew CLI may expose an earlier catalog:

macparakeet-cli spec --json
macparakeet-cli health --json

Transcribe files, folders, media URLs, or podcasts:

macparakeet-cli transcribe /path/to/audio.mp3
macparakeet-cli transcribe /path/to/audio.mp3 --format transcript --no-history
macparakeet-cli transcribe lecture1.m4a lecture2.m4a --output-dir Transcripts --format transcript
macparakeet-cli transcribe --podcast "Lex Fridman episode 400" --format json
macparakeet-cli transcribe /path/to/meeting.m4a --engine nemotron --language auto --format json
macparakeet-cli transcribe /path/to/japanese.m4a --engine cohere --language ja --format json
macparakeet-cli transcribe /path/to/korean.mp3 --engine whisper --language ko --format json

Manage local models and shared app/CLI defaults:

macparakeet-cli models download nemotron-multilingual-1120ms
macparakeet-cli models download cohere-transcribe
macparakeet-cli models download whisper-large-v3-v20240930-turbo-632MB
macparakeet-cli models list
macparakeet-cli models select parakeet-v3
macparakeet-cli config set parakeet-model v2
macparakeet-cli models status

Manage the same prompts used by the app, including collections and versioned inference settings:

macparakeet-cli prompts list --json
macparakeet-cli prompts collections list --json
macparakeet-cli prompts collections add --name "Customer meetings" --json
macparakeet-cli prompts set "Summary" --collection <collection-uuid> --json
macparakeet-cli prompts set "Summary" --temperature 0.3 --max-tokens 2048 --json
macparakeet-cli quick-prompts list --json
macparakeet-cli transforms list --json

Collections organize prompts; recording labels classify recordings and can control prompt availability. Live Ask quick prompts ask questions during a meeting, transcript prompts generate saved outputs, and Transforms rewrite selected text. See the agent guide for editing, version history, collection ordering, and isolated-database examples.

Inspect and update local history, saved meetings, and agent-readable artifacts:

macparakeet-cli history transcriptions --json
macparakeet-cli search '"design review"' --json
macparakeet-cli transcript <id> --around-seq 18 --context 2 --json
macparakeet-cli cards list --source meeting --json
macparakeet-cli retranscribe <id-or-prefix-or-title> --update --json
macparakeet-cli meetings list --json
macparakeet-cli meetings import /path/to/historical-meeting.m4a --started-at 2026-05-14 --json
macparakeet-cli meetings show <meeting-id> --json
macparakeet-cli meetings artifact <meeting-id> --json
macparakeet-cli meetings export <meeting-id> --stdout --format json

Use --format transcript for transcript-only stdout in shell pipelines. Add --no-history when you want a one-off transcription without saving a completed row to MacParakeet history. Multiple inputs or --output-dir write one transcript file per input. models list and models select inspect or update the shared speech default used by the app and --engine app-default; Parakeet rows are parakeet-v3, parakeet-v2, and parakeet-unified, Nemotron rows are nemotron-multilingual-1120ms and nemotron-english-1120ms, Cohere is cohere-transcribe, and Whisper rows use the configured whisper-* model id. The Nemotron, Cohere, and Whisper CLI commands above require their local models to be downloaded first. When developing from source, prefix the same commands with swift run.

Tech stack

Vocabulary

The Vocabulary panel controls how dictated text is cleaned up before pasting. No AI involved — it's a fast, deterministic pipeline that runs in under 1ms.

You choose between two processing modes:

• Raw (default) — Skip cleanup; configured terminal Voice Return triggers are still extracted
• Clean — Run the text through a multi-step pipeline before pasting

The Clean pipeline applies these steps in order:

1. Filler removal — Always strips "uh", "umm", and "uhh". Also strips standalone "um" by default (English hesitation). Turn Also remove “um” off in Vocabulary if you dictate Portuguese or German, where um is a real word.
2. Custom words — Applies your word replacement rules (e.g., "aye pee eye" becomes "API", or "kubernetes" gets capitalized to "Kubernetes"). Case-insensitive, whole-word matching. Words can be toggled on/off without deleting.
3. Voice Return — If you've defined one or more trigger phrases (e.g., "press return" or "zatwierdź") and speak one at the end of a dictation, it's stripped from the output and a Return keypress is simulated after paste
4. Snippet expansion — Replaces short trigger phrases with longer text (e.g., "my signature" expands to "Best regards, David"). Triggers are natural language phrases because that's what the speech engine outputs. Matched longest-first to prevent collisions.
5. Whitespace cleanup — Collapses spaces, fixes punctuation spacing, capitalizes the first letter

When dictation history is enabled, saved entries retain the raw transcript and any processed result so you can inspect what changed.

AI Features

AI features are entirely opt-in and separate from speech recognition — transcription is always local. Configured providers receive text context (transcripts, notes, selected text, or chat as required by the feature), never captured audio.

What it does:

• Summarize — After a transcription finishes, click Summarize and pick a prompt ("Summary", "Action Items & Decisions", "Chapter Breakdown", etc.) or write your own. The LLM processes the transcript and streams back a summary. You can generate multiple summaries per transcript, each in its own tab. Prompts marked as auto-run generate summaries automatically for new transcriptions.
• Chat — Ask questions about a transcript in a multi-turn chat interface. The LLM answers based on the transcript content.
• AI formatter — Optionally run your dictation and file transcripts through your AI provider to clean up grammar, punctuation, and paragraphing. Toggle on/off, customize the prompt, or reset to default.
• Transforms — Select text in any app and press a bound Transform hotkey, such as Control-Option-1 for Polish, to rewrite the selection through your configured LLM provider.

Supported providers:

Setup: In Settings → AI Provider, pick a provider, enter an API key (cloud) or confirm the local server/CLI command is available, select a model, and hit Test Connection. Cloud providers store keys in the macOS Keychain. Ollama and LM Studio can keep LLM inference on-device. CLI subprocess providers run the configured command locally, but that command may contact its own cloud service.

Privacy

All speech recognition runs locally. Parakeet uses the Neural Engine; optional Nemotron Beta, Cohere Transcribe, and WhisperKit engines also run on-device. Your audio never leaves your Mac.

• No cloud STT. The model runs on-device. No audio is transmitted.
• No accounts. No login, no email, no registration.
• Opt-out telemetry. Non-identifying usage analytics and crash reporting go to a self-hosted endpoint only when telemetry is enabled. No persistent IDs, no IP storage, and no transcript/audio content is transmitted. Source code is right here — verify it yourself.
• User-controlled retention. Temporary working audio is cleaned up by its owning flow. Saved dictation, file/media, and meeting audio follows the relevant storage setting; meeting audio is kept by default. Deleting audio does not imply deleting the transcript.

What does use the network: Configured LLM features (including opted-in automatic prompts, formatter, titles, and knowledge cards) send text context to the chosen provider, or whatever service a configured CLI tool uses. Model setup downloads required assets; Sparkle checks for app updates. Media imports use yt-dlp, the public iTunes directory/RSS feeds, and episode downloads. Telemetry/crash reports go to our self-hosted endpoint unless you opt out. Discover independently requests https://macparakeet.com/api/discover.json at app launch by default, even when telemetry is disabled and the Discover page is unopened; it falls back to cached/bundled content offline. Turn off Show Discover in the sidebar in Settings → System → Appearance to hide Discover, cancel its pending feed requests, and stop new feed loads until re-enabled. Explicit feedback and Discover thought submissions also use the network. Core capture and local transcription work offline after model setup, but neither the Discover nor telemetry preference is a global no-network toggle.

Note: Builds from source also send telemetry by default. Opt out in Settings or set MACPARAKEET_TELEMETRY_URL to override.

Contributing

• Report bugs — Open an issue with steps to reproduce and relevant logs or screenshots.
• Discuss new work first — For features or behavior changes, open an issue before starting a PR so we can agree on scope and product fit.
• Submit scoped PRs — Once the issue direction is clear, fork, make the scoped changes, run swift test, and link the issue in the PR.
• Read the specs — Architecture decisions and feature specs live in spec/
• Using a coding agent? Point it at AGENTS.md — the canonical build/test commands, code style, repo conventions, and links to deeper context for Claude Code, Codex, and friends.

Support

MacParakeet is free and open source. If it's useful to you, consider sponsoring.

License

GPL-3.0. Free software. Full license.

Optional Orukeet Core ML preview

Settings → Speech Engines offers Orukeet (preview) among the Parakeet variants. This local adaptation of Parakeet v3 supports 25 languages and leaves the default model unchanged. The 445 MiB download comes directly from Hugging Face, is verified against a pinned integrity manifest and SHA-256, and is compiled for your Mac. Cached transcription needs no network connection. Weights are licensed CC BY-SA 4.0. Preview quality should be evaluated on your own recordings.

The CLI can download it with macparakeet-cli models download parakeet-orukeet, then use it for one recording with macparakeet-cli transcribe --parakeet-model orukeet recording.wav. Native streaming, tail-window dictation preview, and recognition-time vocabulary boosting are disabled for this preview.