Parakeet (Local)

State-of-the-art local transcription with NVIDIA Parakeet

“og:title”: “Stenox Docs”

This is a TRANSCRIPTION provider. It converts your speech to text. For AI enhancement like grammar correction, see AI Enhancement Providers.

Overview

Parakeet is NVIDIA’s state-of-the-art speech recognition model, optimized for Apple Silicon via CoreML. It achieves significantly better accuracy than Whisper while running locally on your Mac’s Neural Engine — no internet required.

Key Benefits:

• Best Accuracy — 6% word error rate (vs Whisper’s 7.4%)
• 100% Private — Audio never leaves your Mac
• Offline — Works without internet connection
• Free — No API keys, no usage costs
• Fast — Optimized for Apple Neural Engine

Privacy & Security

Parakeet processes everything locally on your Mac:

• Audio is captured and transcribed on-device
• No data sent to cloud services
• No internet connection required
• No API keys or accounts needed
• Perfect for sensitive or confidential content

Parakeet is ideal for anyone who wants the best transcription accuracy without compromising privacy.

No API Key Required

Like WhisperKit, Parakeet requires no setup:

1. Select Parakeet in Stenox Settings
2. Download your preferred model
3. Start dictating immediately

No registration, no API keys, no authentication.

Available Models

Model	Size	Speed	Accuracy	Languages
Parakeet v2	~2.5 GB	Very Fast	Best	English only
Parakeet v3	~2.5 GB	Very Fast	Best	25+ languages

Recommended: Parakeet v2 for English-only use. Parakeet v3 if you need multilingual support.

Model Selection Guide

Parakeet v2 - Best for English

Best for:

• English-only dictation
• Highest accuracy available
• Professional writing and technical content

NVIDIA’s TDT (Token-and-Duration Transducer) architecture provides state-of-the-art English transcription.

Parakeet v3 - Multilingual

Best for:

• Non-English languages (25+ supported)
• Multilingual dictation
• Same accuracy as v2 with language flexibility

Supports languages including Spanish, French, German, Chinese, Japanese, Korean, Hindi, and more.

Setup Instructions

1. Open Stenox Settings

   Click the Stenox icon in your menu bar and select Settings.

2. Navigate to Models tab

   Go to the Models tab in the Settings window.

3. Select Parakeet

   Under Transcription Provider, select Parakeet from the dropdown.

4. Choose a model

   Select your preferred model:

   • Parakeet v2 (recommended for English)
   • Parakeet v3 (multilingual)
5. Download the model

   Click Download and wait for the model to download (~2.5 GB).

   The model is stored locally and optimized for CoreML on first use.

6. Start dictating

   Once download completes, you can start using Parakeet immediately!

Performance

Parakeet runs on your Mac’s Apple Neural Engine for fast, efficient processing:

• Processing speed: ~190x real-time on Apple Silicon (M-series chips)
• Typical latency: 2-3 seconds for 10 seconds of audio
• First use: May take a minute to compile the CoreML model (one-time)

Apple Silicon required. Parakeet uses CoreML and the Apple Neural Engine, which are only available on M1 and later Macs.

Storage Requirements

• Model size: ~2.5 GB per model
• Location: Managed by FluidAudio framework

You can delete models you’re not using to free up space via the Models tab in Settings.

Parakeet vs WhisperKit

Feature	Parakeet	WhisperKit
Accuracy	Best (6% WER)	Good (7.4% WER)
Speed	Very Fast	Fast
Model Size	~2.5 GB	142 MB - 483 MB
Languages	English + 25+	100+
Intel Mac	No	Yes
Neural Engine	Yes (CoreML)	Yes (CoreML)

Choose Parakeet when you want the best accuracy and have an Apple Silicon Mac. Choose WhisperKit when you need smaller models, Intel Mac support, or 100+ language coverage.

When to Use Parakeet

Best local accuracy

State-of-the-art transcription running entirely on your Mac.

Privacy is critical

Healthcare, legal, financial, or any sensitive content.

Working offline

Airplanes, remote locations, or unstable internet.

Professional use

High-accuracy transcription for professional writing and documentation.

Troubleshooting

Model download fails

• Check your internet connection
• Ensure you have at least 3 GB of free disk space
• Try downloading again — downloads support resuming

First transcription is slow

• The CoreML model needs to compile on first use — this is a one-time process
• Subsequent transcriptions will be much faster
• You’ll see a “Compiling model” toast while this happens

Wrong language being transcribed

• For English, use Parakeet v2 (English-optimized)
• For other languages, use Parakeet v3 and set language in Profile settings

Poor accuracy for technical terms

• Add terms to Vocabulary for custom replacements
• Parakeet v2 generally handles technical terms well for English

Next Steps

Add AI Enhancement

Use MLX (local) for completely private grammar correction and formatting.

Create Profiles

Set up different profiles for different use cases.

Compare with WhisperKit

See how WhisperKit compares for your needs.

Add Vocabulary

Add custom words and technical terms for better accuracy.