Parlin: Private Mac Transcription and Dictation
Dictate into any app, record meetings with live transcripts and suggested questions, translate as you speak. Parakeet, Whisper, and a bundled local LLM run on-device so audio never leaves your Mac. Submitted to the App Store; direct download lives at getparlin.com.
App Screenshots
Project Overview
Parlin is a native macOS app that turns audio into clean, timestamped text without ever touching the cloud. English dictation runs on NVIDIA Parakeet at roughly a tenth of a second per utterance with about 98% word accuracy on clear speech (1.69% WER on LibriSpeech test-clean); Whisper covers the other 98 languages through WhisperKit, 99 in total. Cleanup and summaries run on a bundled local LLM through the Apple MLX framework, fast even on an M1. On stronger hardware, swap in any local model (Qwen, Llama, Gemma, Mistral): because everything runs locally, the app gets better every time the hardware does.
Meetings are a first-class feature. Record with system audio and watch the live transcript while Parlin suggests questions worth asking; on stop it writes a summary, key points, and action items, editable in place and exportable as a beautifully formatted PDF or Markdown file.
Three input paths cover almost every use case. Drop an MP4, MOV, MP3, WAV, M4A, FLAC, or WebM file straight onto the window. Paste a media URL and the app downloads, normalizes, and transcribes it. Or hold ⌃⌥Space anywhere on the system to record dictation that pastes into the focused app the moment you let go.
A picker covers Whisper Tiny through optimized Large v3 Turbo so the user trades speed for accuracy on their own terms. Auto language detection covers 16 languages out of the box, and picking an output language turns on near real-time translation: speak one language, the text lands in another. Output exports as plain text, SRT, or VTT, with or without word-level timestamps.
The codebase ships as two SKUs from one source. The App Store SKU runs sandboxed and supports direct media URLs only. The direct-download SKU at getparlin.com bundles broader URL support via yt-dlp for users who need it. Same engine, different distribution paths.
Key Features
- Fully offline: Whisper and the cleanup LLM both run on-device
- 99 languages: Parakeet for English (about 98% accuracy on clear speech), Whisper for the rest
- Meeting recording with live transcript, summary, key points, and action items
- Live suggested questions while the meeting runs
- Near real-time translation: dictate in one language, output in another
- Meeting notes export as beautifully formatted PDF or Markdown
- Global hold-to-talk dictation (⌃⌥Space) that pastes into any app
- Drag-and-drop transcription for MP4, MOV, MP3, WAV, M4A, FLAC, WebM
- Media URL ingest with yt-dlp on the direct-download build
- Whisper model picker from Tiny (77 MB) through optimized Large v3 Turbo
- Bundled MLX cleanup LLM runs on any M-series Mac; swap in any local model via Ollama
- Word-level timestamped export to TXT, SRT, or VTT
- Sandboxed App Store SKU and notarized direct DMG from a single codebase
Tech Stack
Platforms
Project Details
2026
2 months
1 engineer
Metrics
On-device. Zero cloud dependency.
Parakeet + WhisperKit on Apple Silicon
Privacy-first transcription and dictation for macOS
Business Impact
Parlin shows DevX Group's ability to ship a polished native macOS app with on-device AI on two distribution channels at once. Same codebase serves the Mac App Store under sandbox constraints and a notarized direct DMG with broader URL support. Demonstrates the engineering judgment to design around App Store guidelines without compromising the product the team actually wanted to build.
Interested in a Similar Project?
Let's discuss how we can build something great for your business.