Technical details · updated 2026-08-19
The exact pieces, for people who check.
The homepage tells you what Parlin does. This page tells you what runs it: the models, the frameworks, and the numbers, so you can verify the on-device claim yourself before you buy.
Speech to text
- English dictation
- Parakeet TDT 0.6B v2 (NVIDIA), on the Neural Engine
- English accuracy
- 98.3% on clear speech: 1.69% word error rate on LibriSpeech test-clean, from the model card's published Open ASR Leaderboard results. Real meetings with far-field mics score lower on the same benchmark suite
- Latency
- About 0.2 seconds from key-up to text
- European languages
- Parakeet TDT 0.6B v3 (NVIDIA), same engine and speed, for Bulgarian, Croatian, Czech, Danish, Dutch, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Russian, Slovak, Slovenian, Spanish, Swedish and Ukrainian
- Other languages
- Whisper via WhisperKit, 70+ more languages including Persian, Arabic, Chinese, Japanese, Korean, Hindi and Turkish
- Higher accuracy
- Larger Whisper models downloadable in Settings
- Files and meetings
- Same engines, plus speaker diarization
Cleanup
- Model
- Qwen3 1.7B, downloaded once and managed by Parlin
- Runtime
- Apple MLX, on the Mac's GPU (Metal)
- Setup
- One automatic download on first use. No account, no API key
- Bring your own
- Direct build can use any Ollama model from Settings
Runtime
- Hardware
- Apple Silicon (M1 or later), macOS 14.6+
- Memory
- About 2.4 GB with the default models loaded
- Network
- No audio or transcripts, ever. The only traffic: model downloads, update checks, and license validation on the Direct build
- Telemetry
- None. No audio, transcripts, or usage analytics leave the Mac
Under the hood
Open where it should be. Closed where it matters.
We don't invent privacy from scratch. We assemble it from tools you can audit yourself, then add the layer that makes them feel like one app on your Mac.
Public projects with permissive licenses. Read the source, run the models yourself, verify the privacy claim end to end.
- ParakeetCC-BY-4.0
NVIDIA's open speech model. Fast dictation in English and 24 European languages on your Mac's Neural Engine, so text lands in about a fifth of a second.
- WhisperMIT
The open speech model behind WhisperKit. Covers the other 70+ languages, files, and meetings, all on your Mac.
- MLXMIT
Apple's machine learning framework for Apple Silicon. Powers the built-in cleanup LLM and runs it on the Mac's GPU.
- Apple SiliconPlatform
Every model path is tuned for the Neural Engine and Metal. M-series Macs only, by design.
The layer that turns four open projects into one product that feels native. Our work, our code, our continued investment.
Developer-vocabulary tuning
A terminology layer for code, CLIs, library and identifier names so paths and camelCase survive transcription without manual fixes.
Push-to-talk pipeline
Sub-second key-down to paste across any focused app. Capture, transcription, cleanup, and insertion stitched into one perceived action.
Cleanup orchestration
Context-aware prompting that fixes grammar, punctuation, and capitalization without rewriting your voice. The model is yours; the conductor is ours.
macOS integration
Accessibility paste, focus tracking, system shortcuts, and a recordings library built directly against Apple frameworks.
Project names and logos are trademarks of their respective owners. Parlin is built with these projects but is not affiliated with or endorsed by them.
You checked. It holds up.
No card for the trial. $39 once, rising to $69 on Oct 31, 2026. No subscription, ever.