How on-device dictation
actually works
Five years ago, good speech recognition meant a data center. Today the same quality runs on the laptop you already own. Here's the pipeline, why it became possible, and what it means for privacy and price.
Microphone to cursor, all on one machine
1. Capture
Press the hotkey (F9 by default) and Blabb starts recording from your microphone. Audio is held in memory on your machine for as long as the session runs — there is no upload step, and nothing is streamed anywhere.
2. Voice activity detection
A small neural voice-activity model (TEN VAD) watches the incoming audio and splits it into speech segments at genuine silence. This is the part that lets you pause, think, and carry on: pauses become segment boundaries, not a reason to stop. Blabb never ends the session on its own — you decide when you're finished.
3. Speech to text
Each segment feeds a neural speech-recognition model running locally. Open-model research made this a solved problem: Cohere Transcribe, the openly licensed model Blabb ships, topped the Hugging Face Open ASR Leaderboard on release at 5.42% word error rate (March 2026) — accuracy you can verify, not take on our word. About 1.5 GB on disk, included in the installer — no download step, no API key.
4. Cleanup — optional
Raw transcription is faithful but messy — fillers, missing punctuation, "twenty three" instead of 23. A compact local LLM (IBM Granite, built under an ISO 42001-certified AI management process) can rewrite it under strict instructions: fix the mess, never invent content. Six styles set how far it goes. A new install starts on Keep original, so this step does nothing until you choose otherwise.
5. Text entry at the cursor
The text is delivered into whatever window has focus, by one of four methods you choose in Settings — direct typing, paste, simulated keystrokes, or UI automation — which is why it works in terminals and clipboard-blocked remote sessions (details). To the receiving app, you just typed very fast. The one place Windows won't allow it is a window running as administrator.
Three shifts made the cloud optional
Quantization
Model weights compressed from 16-bit to ~4-bit shrink memory and compute several-fold with minimal quality loss — the difference between "needs a server GPU" and "runs beside your browser." Both of Blabb's models ship quantized; inference uses your GPU when present and falls back to CPU when not.
Open model releases
Openly licensed speech and language models now match or beat proprietary cloud APIs on public benchmarks — verifiable on the Open ASR Leaderboard rather than taken on a vendor's word. Open licensing is also why models are swappable: when a better one ships, an update brings it to you.
The economics follow
Cloud dictation pays for GPU time on every word and bills you monthly for it — $10–20/month is the going rate. On-device inference uses hardware you already own, so there's no marginal cost per word. That's the entire reason Blabb can be USD $4.99/month, unlimited: there's no inference bill to pass on. And privacy isn't a policy promise — your audio and transcripts are never uploaded, so there's no store of them to leak.
Runs on your PC — so speed follows your hardware
Because nothing runs in the cloud, the machine doing the work is yours. On a capable PC it feels instant; on modest hardware there's a real, honest wait — here's what to expect.
Runs on any Windows 11 PC
Windows 11 (x64) and a microphone are all it needs — no GPU, no NPU, no internet. Both models ship in the installer and run on your CPU. It will work on a modest laptop; the question is how fast, not whether.
What a modest machine feels like
On a low-end setup — 4 CPU cores, 8 GB of RAM, no GPU — expect a short sentence to come back in a second or two, with AI cleanup adding more on top. We publish that as an estimate, not a benchmark: we haven't measured a figure we'd be willing to print as one, and your hardware, microphone and settings all move it. The very first dictation after launch is slower while the models load into memory; after that warm-up, every dictation is quicker.
Low-spec? Transcribe-only is the light path
The AI cleanup is what's heavy — not the transcription. The speech model (Cohere Transcribe) is highly accurate and does most of the work on its own, so leaving cleanup on Keep original — which is where a new install starts — skips the LLM entirely and is the fastest way to run Blabb on a slow machine. Cleanup is the optional polish on top.
Recommended for the smoothest ride
For fast results with cleanup on, we suggest a dedicated GPU, 16 GB or more of RAM, and a few extra CPU cores. Both stages use a graphics card when they can, with nothing to install: speech recognition picks the fastest backend its runtime finds — Vulkan on NVIDIA, AMD, and Intel, then CPU — while cleanup uses LM-Kit’s own layer offload, which sizes itself to your available video memory and drops back to the processor when there isn't enough. All of that happens automatically — Blabb offloads to the graphics card when it can and falls back to the processor when it can't, with nothing for you to configure.
A note on processor type. Blabb is a native x64 app, built for Intel or AMD PCs. Windows-on-Arm machines — the Snapdragon-powered Copilot+ PCs, including the Surface Laptop 7 — can only run it through x64 emulation, where the on-device AI is much slower. Choosing a Microsoft Surface for Blabb? Pick an Intel model, such as the Surface Laptop 6 for Business; native Arm support isn't available yet.
Everything that touches the network
Privacy by policy asks you to trust a promise. Privacy by architecture is checkable — here is the complete list, verifiable by anyone with a network monitor.
The licence check
Blabb asks Polar, the payment provider, whether your licence key is valid. No audio, no transcripts, no usage — just the key. Once verified, Blabb keeps dictating for 30 days offline before it needs to ask again.
Crash reports — only if you opt in
Off by default. If you switch it on in Settings, crash reports go out with dictation content scrubbed. Leave it off and this call never happens.
Model downloads — only if you change models
The standard models ship in the installer. Choose a different model in Settings and that one is downloaded once, then cached. Never your data going up — only fixed weights coming down.
That's the whole list
No analytics, no usage tracking, no telemetry, no account. Your dictation history is encrypted on your PC with Windows DPAPI and never leaves it. The cable test: pull the network cable mid-sentence — dictation doesn't notice.
Data-center dictation quality. No data center.
Two weeks free, models in the installer, dictation that runs on your PC.
FREE TIER 2,000 WORDS/DAY · 14-DAY UNLIMITED TRIAL IN-APP · CANCEL ANY TIME