Kalam — Private On-Device Dictation

Hold Ctrl+Space, speak, release. Your words appear where you were typing. 100% on-device — no audio ever leaves your computer.

Hold Ctrl+Space. Speak. Release. Your words appear exactly where you were typing. Kalam is push-to-talk dictation that runs speech recognition entirely on your own computer. No cloud service, no account, no API key, no subscription. Your microphone audio and your transcripts never leave your device — turn on airplane mode and dictation keeps working. HOW IT WORKS • Click into any text field on any website • Hold Ctrl+Space and speak (also configurable: Ctrl+Shift+Space, Right Ctrl, F8 — and F9 always works as a fallback) • Release the key — your transcript is inserted at the cursor • No text field focused? The transcript goes to your clipboard instead WHY ON-DEVICE MATTERS Most dictation tools stream your voice to a server. Kalam does not have a server. Speech models run inside the extension using WebGPU (with a multi-threaded CPU fallback), so: • Your voice is never uploaded, never stored, never used to train anything • It works on flights, in dead zones, and on locked-down networks • There is nothing to subscribe to and no usage quota • You can verify all of it — Kalam is open source under the MIT license CHOOSE YOUR ENGINE Download only what you want, and delete any of it from the Models tab: • Cohere Transcribe (~1.2 GB) — top accuracy, 14 languages including Arabic; needs a GPU • Whisper Large v3 Turbo (~600 MB) — around 100 languages • Whisper Base (~80 MB) — light and CPU-friendly, the fastest way to get started • Parakeet CTC 0.6B — English only, very fast, architecturally unable to hallucinate text • Gemma 4 E2B (~3 GB) — experimental engine, and optional transcript polish for spoken punctuation and cleanup "Auto" measures your hardware and picks for you, downgrading automatically if a model turns out to be too slow on your machine. PRIVACY, VERIFIABLY • The speech runtime and its WebAssembly binaries ship inside the extension — no CDN calls at runtime • The only network request Kalam ever makes is the one-time model download from huggingface.co. After that you can go offline permanently • The microphone is opened only while you hold the key and released the instant you let go — Chrome's own recording indicator confirms it • No analytics, no telemetry, no crash reporting, no ads, no identifiers • Web pages get nothing: they cannot reach your microphone through Kalam, read its status, or spoof its interface • Full source: https://github.com/madforstrength/kalam SETUP 1. Install, then click the Kalam toolbar icon and allow microphone access (choose the persistent "while visiting this site" option) 2. Open the Models tab and download an engine — Whisper Base is the quick start, Cohere Transcribe is the accuracy pick on a GPU 3. Click into any text field, hold Ctrl+Space, and speak Requires Chrome 124 or later. Works best where Chrome can use your GPU (check chrome://gpu for WebGPU); otherwise it falls back to multi-threaded CPU inference, which is slower but works. BETA Kalam works end-to-end and has been hardened against a long list of real-world failures, but it has been tested on a limited range of hardware. If something misbehaves on your machine, please open an issue with your OS and GPU — that feedback is exactly what this stage is for. Kalam (Arabic: كلام) means "speech".