2-way conversations drive
parallel teams of agents
with only your voice.
Speech Layer is local-first speech for Linux and macOS. Your words are transcribed on your own machine, routed to AI agents that work in parallel, and when they finish your desktop speaks the answer back. Private by design, hands free.
Hear a real conversation with your computer
Press play and listen: Speech Layer hears the request, puts agents to work on it, often several at once, and speaks the answer back. Six short examples below.
Code by voice. Agents fan out and report back.
Speak a task mid-flow: it is transcribed on your own machine, fanned out to coding agents that work in parallel, and when they finish your machine reads the summary back. Hands free, private by design.
Speech in, action out
A thin, dependable layer between your voice and your toolchain. No cloud round trips required.
Always-ready wake word
Say the wake word mid-flow and start dictating: no window switching, no hands off the home row.
Local speech-to-text
Whisper-class transcription runs on your machine. Dictate a commit message or a whole feature brief, online or off.
Agents that code
Intents fan out to scripts and coding agents that do the work: dispatch a fix, run the tests, review a diff. Mail and browsing come along too.
Spoken confirmation
Fast local text-to-speech reads back summaries and asks the next question, so you steer a whole fleet of workers without touching the keyboard.
Built like infrastructure, not a demo
Speech Layer is the voice plumbing your desktop has been missing: boring in the best way, and always there.
Private by design
Audio capture, wake word detection, and transcription happen on your hardware. Your code and your voice never leave the machine unless you ask them to.
Local-first
Core features keep working with the network cable pulled. Cloud models are an option you opt into, never a requirement.
Linux and macOS native
First-class support for real desktops: PulseAudio and PipeWire on Linux, Core Audio on macOS, sensible packaging on both.
Scriptable and agent-ready
Every intent is just a command. Wire voice to git, your test runner, your multiplexer, or full coding agents with a few lines of config.
Runs where you work
Speech Layer targets Linux desktops and macOS, with packages and docs for the distributions people actually use.
Linuxkernel 5.x and newer
macOSApple Silicon and Intel
Ubuntu22.04 LTS and newer
Fedoracurrent releases
Debian12 bookworm and newer
Arch Linuxrepos and AUR
Artixrunit and OpenRC
NixOSflake providedAlso tested on openSUSE Tumbleweed. Other systemd or OpenRC based distributions generally work via the generic installer.
One loop, five stages
Each stage is a separate, replaceable component. Swap the transcriber, the intent router, or the voice, the pipeline stays the same.
Wake word
Low-power detector listens for your chosen phrase and opens the mic only then.
Speech-to-text
On-device transcription turns your request into text in well under a second.
Intent and agent
The router matches what you said to a script, tool, or AI agent with the right context.
Action
The agent does the work: reads mail, moves windows, runs builds, controls the desktop.
Spoken reply
A fast local voice confirms what happened, so you never break flow to check a screen.
Up and talking in minutes
One command installs the daemon, a default wake word, and a starter set of intents. Point it at your own scripts whenever you are ready.
# Linux and macOS $ curl -fsSL https://speechlayer.com/install.sh | sh # check the daemon $ speechlayer status # say the wake word and go $ speechlayer start
A small stack of dependable open parts, all running on your machine.