Press a key, speak, and your words appear. Wayland‑native. Fully configurable. Local or cloud models. Built in Go because concurrency is fun—and useful.
A tiny story about a keypress
The first time Hyprvoice felt “right,” I was staring at a blinking cursor. I tapped Super+R, said “meet at three tomorrow,” tapped Super+R again, and watched the sentence land exactly where my cursor was. No overlay, no floating window, no mouse. Just a gesture.
That single gesture is the whole point. Voice is incredible when it disappears into your flow instead of dragging you into a new app. Hyprvoice tries to stay invisible.
Why I built this (and why I consider it my best open‑source project)
Wayland desktops—and Hyprland in particular—deserve voice‑to‑text that isn’t a museum of X11 hacks. I couldn’t find anything that was both Wayland‑native and fully configurable for Linux nerds who want control: local models for privacy, cloud models for convenience, and enough knobs to tune behavior without recompiling. So I wrote it.
Selfishly, I also wanted an excuse to go deep on Go: goroutines, synchronization, timeouts, and clean IPC. Hyprvoice became a playground where those ideas turned into a tool I actually use every day.
The UX is the product
Hyprvoice works like a camera shutter. Tap to start recording; tap to stop and inject. Desktop notifications tell you what’s happening—recording, transcribing, injected—but the UI never gets in your way. If the compositor/app supports direct typing, Hyprvoice types for you. If not, it quietly falls back to the clipboard (and even restores what was there before).
If you try one thing, try this:
# ~/.config/hypr/hyprland.conf
bind = SUPER, R, exec, hyprvoice toggle
Open an editor, a terminal, a browser—tap, speak, tap. That’s the demo.
How the architecture clicks (without drowning you in plumbing)
There’s a tiny daemon with a Unix socket. The CLI/keybind sends one‑byte commands to it. Inside, a pipeline moves through clear states: idle → recording → transcribing → injecting → idle. PipeWire captures audio, a transcriber turns it into text, and an injector delivers it where your cursor lives. Everything is guarded by timeouts so no part can get stuck.
Most days you never notice any of this. It’s the point.
Wayland‑native or bust
Hyprvoice is built for Wayland compositors. No X11 detours, no brittle bridges. Direct typing via wtype feels magical when it works; when it doesn’t, the clipboard path makes sure your words still arrive. Either way, you stay in the same app, in the same flow.
Recording that respects your CPU (and your attention)
Speech doesn’t need 96kHz stereo. Hyprvoice records at 16 kHz mono s16: exactly what speech models expect, easy on bandwidth and battery, and fast to ship through the pipeline. Sessions are bounded (e.g., a 5‑minute timeout) so you don’t end up recording your entire afternoon by accident.
Transcription is a plug
Today, Hyprvoice talks to OpenAI Whisper (online). Next up: whisper.cpp for fully local models you can grab from Hugging Face. After that: streaming transcription, so text appears while you speak. The design that makes this possible is a tiny interface—change one line in the config to swap engines.
// One skinny seam for engines
// internal/transcriber/transcriber.go
type TranscriptionAdapter interface {
Transcribe(ctx context.Context, audio []byte) (string, error)
}
# ~/.config/hyprvoice/config.toml
[transcription]
provider = "openai" # soon: "whisper_cpp" for local/offline
model = "whisper-1"
language = "" # "" = auto-detect
# api_key can come from env: OPENAI_API_KEY
Adapters make Hyprvoice extensible without turning it into a framework—just enough structure to keep new providers tidy.
Config you can tweak mid‑flow
The daemon watches its TOML config and hot‑reloads it. Want to lengthen the typing timeout or switch injection mode? Save the file and try again—no restarts, no flags, no yak‑shaving.
[injection]
mode = "fallback" # "type" | "clipboard" | "fallback"
restore_clipboard = true
wtype_timeout = "5s"
clipboard_timeout = "3s"
It’s a small thing that makes daily life nicer.
When things go wrong, make it boring
Software fails. What matters is how. Hyprvoice cleans up stale sockets/PIDs on start. Every external tool has a timeout. If direct typing can’t reach an app, the clipboard path quietly takes over and your original clipboard is restored in the background. The log tells the story; you keep typing.
And the secret sauce behind that calm? Go’s context everywhere. Every long‑running operation and external call carries a deadline or cancel. Toggle again and the entire pipeline unwinds. A timeout fires and recording stops; sockets close; goroutines exit. On shutdown, WaitGroups drain, contexts propagate, and the system cleans up predictably.
Here’s the pattern you’ll find all over the codebase:
runCtx, cancel := context.WithTimeout(ctx, p.config.Recording.Timeout)
p.setCancel(cancel)
p.wg.Add(1)
go p.run(runCtx)
That one pattern makes the whole system resilient—nothing hangs forever.
Privacy dials you actually control
You choose where your audio goes.
Prefer convenience? Use the cloud provider and you’re a config line away from high‑quality transcription. Prefer privacy or hacking on models? Switch to whisper.cpp and keep everything on your machine. Hyprvoice doesn’t moralize; it gives you the dial.
Try it
On Arch:
yay -S hyprvoice-bin # or yay
hyprvoice install
Bind the key, speak, and see if it changes your habits. If you’re on another distro, grab the release binary or build from source.
What I’m doing next
Streaming transcription is the next big step—seeing partial text while you speak makes voice input feel instant. I’m also building the whisper.cpp adapter and exploring per‑app rules (for example, automatically disabling injection in password fields).
To grow the project, I’ll share and sponsor posts on X/Twitter, Hacker News, and Reddit to recruit testers and contributors.
Want to help?
If you’re new to open source, Hyprvoice is a friendly place to start. Try it on your setup and open an issue with logs if anything feels off. If you’re a builder, adapters and tests are the best leverage. If you’re just curious, share latency and accuracy numbers from your machine—those reports help guide defaults.
Hyprvoice tries to be invisible. The best compliment would be that you forgot it was running and kept talking.