Browser dictation to the focused window
Speak into the browser and have the selected client type into the application you're using.
VOICE KEYBOARD / SELF-HOSTED DICTATION
Speak here. Type where you work.
Dictate from the browser, transcribe with Whisper on your own server, and send text to the selected desktop or Bluetooth keyboard client.
Public source. Self-host on your infrastructure.
Built around your workflow
Cross-device dictationSelf-hosted speech workflowsOptional Bluetooth keyboard setups
Inside Voice Keyboard
Your voice becomes input.
Self-hosted Whisper transcription
Run Whisper on your own GPU, with a slower CPU option when a compatible GPU is unavailable.
Native desktop clients
Use a small background client on Windows, macOS, or Linux, then choose the machine that should receive your text.
Optional ESP32 Bluetooth keyboard
Pair an ESP32 or ESP32-C3 as a Bluetooth keyboard when you want a receiver without installing a desktop client on the target device.
Languages, translation and Hinglish
Choose a transcription language, use the English-translation mode, or write Hindi speech in Latin letters with Hinglish.
Address resilience and device selection
Keep local and remote addresses configured for the same server, and let supported clients reconnect to an available route.
Inside the pipeline
Continuous audio. Deliberate text delivery.
Voice Keyboard / system flow01 → 06
- 01
Capture
Browser microphone + AudioWorklet
- 02
Transport
16 kHz mono signed PCM over WebSocket
- 03
Segment
Silero VAD selects utterance boundaries
- 04
Recognize
Queued WAV requests to shared Whisper model
- 05
Join
Context, spaces and seam punctuation
- 06
Deliver
Selected native or Bluetooth receiver
01 / MECHANISM
Small audio frames, meaningful speech boundaries
An AudioWorklet converts microphone samples into 16 kHz mono PCM and sends 1,600-sample transport chunks. Silero examines 32 ms windows and uses speech hysteresis, pause detection and bounded utterance length to choose recognition boundaries.
02 / MECHANISM
Whisper recognizes phrases while you keep speaking
Each completed utterance becomes an ordinary WAV transcription request. A shared inference semaphore controls model concurrency. This is phrase-level streaming, rather than Whisper token streaming or speculative partial text.
03 / MECHANISM
Sentence continuity belongs in the joiner
The next request receives recent raw text as context. A seam-aware joiner handles spaces, ellipses and punctuation at forced cuts, then sends appendable segments to the account’s selected receiver. It does not rewrite words already typed.
04 / MECHANISM
A microphone and a practical control surface
Desktop agents or supported ESP32 Bluetooth receivers turn text into input. The browser keyboard and touchpad provide a separate path for manual keys, pointer movement and scrolling. Receiver permissions and script support depend on the platform.
Implementation parameters · not benchmarks
- Audio frame
- 1,600 samples
- transport default
- VAD window
- 32 ms
- analysis window
- Inference slots
- 1
- configured default
Continuous audio capture and utterance-level recognition run as separate stages. The joiner prepares recognized segments for the selected receiver; the engineering article explains queue and reconnect behavior.
Read the full engineering storyHow it works
A clear path from
setup to useful work.
- 01Open the browser microphone
- 02Transcribe on your Whisper server
- 03Type through your selected receiver
A closer look
One receiver, chosen by you.
Select the desktop that should receive your words. Or pair a supported ESP32 Bluetooth keyboard when the target device should not need a desktop client.
/ workflow
Browser microphone → Whisper → Selected client → Focused app
Conceptual workflow · configured for your environment
Practical questions
Know what
fits your setup.
Where does speech recognition run?
Whisper runs on your configured server. GPU and slower CPU options are available. Initial model downloads, server resources and configured connectivity are required; speed and recognition quality depend on your setup.
Can it type into any application?
The selected client types into the focused window and needs platform permissions. macOS needs Accessibility access. A non-elevated Windows client cannot type into elevated applications. Linux Wayland needs uinput setup and has US-layout ASCII limitations.
How does the Bluetooth option work?
Supported ESP32 or ESP32-C3 firmware connects over Wi-Fi to the server and presents a Bluetooth keyboard to the target. It needs power, pairing and network access. US keyboard mapping normalizes accented Latin text and skips unsupported scripts.
What do I need in the browser?
Microphone access requires trusted HTTPS and microphone permission. You choose one active receiver. Transcription language, English translation and Hinglish options still depend on the receiver’s script support.
Make it work for you