Offline-First Architecture

Your voice never
leaves your Mac.

Every word you speak is processed entirely on your device by Voxtral AI from Mistral — an open-source model running on Apple Silicon. No cloud servers. No API calls. No data collection. Ever.

Your voice → Internet → Cloud server → Response → Your appCloud dictation
Your voice → Your Mac → TextEdgeWhisper

Why offline-first?

Cloud-based dictation tools send your raw audio to remote servers for processing. That audio can contain passwords, financial data, medical information, trade secrets — everything you say. We chose a fundamentally different architecture.

Architectural Privacy

It's not a policy choice — it's a design guarantee. EdgeWhisper physically cannot send your data anywhere because the AI runs entirely on-device. There is no server to send data to.

Works Without Internet

After the one-time AI model download, EdgeWhisper works completely offline. Airplane mode, coffee shop without WiFi, classified facility — dictate anywhere, anytime.

Zero Data Collection

No audio recordings. No transcription logs. No usage analytics. No telemetry. No cookies. No device fingerprints. We engineered the app to collect absolutely nothing.

The problem with cloud dictation

Every cloud-based dictation tool shares the same fundamental flaw: your voice has to leave your device.

RISK 01

Your audio is on someone else's server

Cloud dictation services receive your raw audio stream. Even with encryption in transit, your voice data is decrypted and processed on servers you don't control.

RISK 02

Data can be retained for training

Many services use your audio to improve their models. Your private conversations, medical notes, and legal documents may become training data.

RISK 03

Vulnerable to breaches

Any data stored on remote servers can be breached. Voice data is uniquely sensitive — it's biometric, containing your identity, emotions, and context.

RISK 04

Requires constant connectivity

No internet? No dictation. Cloud services fail when you need them most — on flights, in remote areas, or in secure facilities.

How EdgeWhisper is different

Privacy isn't a feature we added. It's the foundation we built on.

AI Model

Proprietary, closed-source model on remote servers

Voxtral Mini 4B from Mistral AI — open-source (Apache 2.0), runs locally

Audio Processing

Audio streamed to cloud, processed on third-party infrastructure

Processed in-memory on Apple Silicon's Neural Engine, never leaves RAM

Network Required

Always — no internet means no dictation

Never — fully offline after one-time model download

Data Collection

Audio, metadata, usage patterns, device identifiers

Nothing — zero telemetry, zero analytics, zero tracking

Account Required

Email, password, often phone verification

None — download and start dictating immediately

Compliance

Depends on vendor's data handling practices

Inherently compliant — no data leaves the device to regulate

Auditability

Closed-source, trust the vendor's claims

Open-source AI model, verifiable architecture, transparent processing

Built for people who care about privacy

If your words matter, they should stay yours.

Healthcare Professionals

Dictate patient notes, clinical observations, and medical records without HIPAA concerns. Nothing leaves your Mac.

Legal Professionals

Draft contracts, case notes, and client communications. Attorney-client privilege stays intact because no third party ever processes your audio.

Developers & Engineers

Dictate code comments, documentation, and Slack messages. No risk of proprietary code or architecture discussions leaking to cloud servers.

Journalists & Writers

Protect your sources. Draft sensitive stories. No cloud service can be subpoenaed for audio that never left your computer.

Business Executives

Dictation during board prep, M&A discussions, and strategy sessions. Your competitive intelligence stays on your device.

Security-Conscious Users

Anyone who values privacy. Passwords, financial discussions, personal reflections — keep it all local.

Technical guarantees

Don't take our word for it. Verify it yourself.

Open-Source AI Model

EdgeWhisper uses Voxtral Mini 4B Realtime from Mistral AI, released under the Apache 2.0 license. The model weights, architecture, and training methodology are publicly available on Hugging Face for anyone to audit.

View model on Hugging Face →

Apple MLX Framework

The model runs via Apple's MLX framework — a machine learning framework for Apple Silicon. This is Apple's own technology, optimized for on-device inference with no cloud component.

Learn about MLX →

macOS App Sandbox

EdgeWhisper runs inside macOS's App Sandbox, restricting what the app can access. It only requests Microphone (for voice capture) and Accessibility (for text injection) — nothing else.

Read our privacy policy →

Zero Network Calls

After the initial model download from Hugging Face, EdgeWhisper makes zero network requests. You can verify this with any network monitoring tool like Little Snitch or Wireshark. Try it — you'll see nothing.

“We didn't build EdgeWhisper to be private — we built it so that privacy is the only possible outcome. When the AI runs entirely on your Mac, there is no server to breach, no data to sell, and no policy to violate.”

Your voice deserves
better than the cloud.

Download EdgeWhisper free. Powered by Voxtral AI from Mistral, running 100% on your Mac. 5 minutes a day, zero setup, total privacy. Join developers and professionals switching from cloud-based dictation.

macOS 14+ · Apple Silicon (M1/M2/M3/M4) · Voxtral Mini 4B · Apache 2.0 AI · ~3 GB model