Inside the macOS app, your audio never leaves your Mac.

Blazing Transcribe the desktop app runs entirely on your machine. Your microphone audio is processed by the Apple Neural Engine and never touches a server. No cloud API. No audio uploads. No transcription text sent anywhere. We built it this way because local processing isn't just more private — it's why it's 155x real-time. Any separate web demos on this site are labeled explicitly if they use temporary server-side processing.

* * * * * * * * * * * * * * * * * * * * * * * *

$ cat AUDIO_PIPELINE.md

How your voice is processed

[ data flow ]

MIC

Microphone capture

Audio is captured and converted to 16kHz mono

on-device

BUF

Ring buffer (in-memory)

Live samples are held in a circular buffer in RAM — processed on-device, never uploaded

on-device

VAD

Voice activity detection (Silero VAD)

ML model detects when you’re speaking — runs locally on CPU

on-device

ANE

Transcription (NVIDIA Parakeet TDT)

Speech-to-text runs on the Apple Neural Engine — no cloud API

on-device

TXT

Text delivery (CGEvent injection)

Text is typed directly into the focused app via macOS keyboard events

on-device

Every step from microphone to typed text runs locally on your Mac. No audio is sent to any server at any point, and no transcription text is transmitted anywhere. Audio lives on your Mac only — held in memory during live capture, with longer recordings and any saved history cached locally — and is never uploaded.

* * * * * * * * * * * * * * * * * * * * * * * *

$ netstat --app blazing-transcribe

What touches the network

Blazing Transcribe makes exactly three types of network calls. None of them involve your audio or transcription text.

What	Why	What's sent	Audio/text sent?
Analytics	Aggregate usage stats	Event names only (e.g. "appLaunched")	Never
Update check	Check for new versions	Current app version	Never

* * * * * * * * * * * * * * * * * * * * * * * *

$ cat ANALYTICS.md

About our analytics

We use TelemetryDeck, a privacy-first analytics provider based in the EU. Here's what they receive:

Event names — things like "appLaunched", "transcriptionCompleted", "recordingStarted". Just the name of the event, so we know which features are used.

Basic counters — word count per transcription, which recording mode you use. Aggregate numbers, nothing identifiable.

What is never sent: your transcription text, audio data, microphone input, file contents, or anything personally identifiable. TelemetryDeck doesn't use cookies, doesn't track across apps, and doesn't sell data.

We use this to understand which features matter so we can make the app better. That's it.

* * * * * * * * * * * * * * * * * * * * * * * *

$ cat VERIFY.md

Don't trust us. Verify.

Most transcription apps say "local processing" but give you no way to prove it. We've published the audio pipeline architecture on GitHub so you can see exactly how your voice is handled.

The repo documents the complete data flow from microphone to typed text, the models used, and every network call the app makes. You can read the architecture and confirm for yourself that no audio or transcription data ever leaves your machine.

View on GitHub →

* * * * * * * * * * * * * * * * * * * * * * * *