Local. Fast. Honestly built.
No marketing trick replaces engineering. Here's what's actually under it.
Local-first audio pipeline
Audio captured through WASAPI, transcribed by faster-whisper running on your GPU (CUDA optional, CPU fine). Nothing leaves your machine. No telemetry, no analytics, no cloud round-trip.
Streaming dictation
Live partial transcriptions update every second in a floating overlay. The final paste uses a Ctrl+Z undo-cycle to swap drafts cleanly — the result of three rounds of multi-model code review.
Hardware integration
Direct ctypes calls into IAudioEndpointVolume for volume control (because pycaw proved unreliable). TCP M-code protocol for FlashForge printers. JSON-RPC for Hyperion LED strips. AHK for Electron apps that ignore Win32.
Plugin system
Drop a Python file in plugins/commands/, decorate with @command, you have a voice command. Ships with 18 plugins — smart home, music, printers, macros, audio switching, screen recording, GIF search, scroll, text marker, volume/mute, Stremio, and Ava (local LLM voice assistant).
Deterministic first. AI when needed.
High-frequency commands like "open Chrome", "lights red", or "pause print" stay instant and predictable. No AI in the loop, no cloud latency, no chance of being misunderstood. Open-ended requests can optionally route to AI systems you configure yourself — local LLMs, your own API keys, whatever you trust. You decide where the boundary lives.
Nothing fancy. Just things that work.
For the technically curious. You don't need to know any of this to use Samsara.
Demo, recorded in the garage.
A real session, real desk, no edits beyond trim. Wake word, lights, apps, dictation, a sketch on the side.
Originally posted to Instagram. Sound on recommended.
Where this is going.
Recently shipped: Voice AI (Ava) with local Ollama, Text-to-Speech with smart audio ducking, Show Numbers overlay for hands-free clicking, Command Cheat Sheet overlay, Mouse 4 walkie-talkie command mode, window manager with saved layouts, Smart Actions webhook bridge, deferred text selection (mark here / select to here), 5-speed scrolling, repeat/again command, volume & mute via Core Audio API, Stremio voice control, single-stream audio fan-out, command packs with per-user toggles, persistent history with search and session grouping.
Coming next: Speaker verification (local voice enrollment via Resemblyzer). Voice Training panel — calibration phrases, corrections dictionary, test mode. End-word and cancel-word dictation protocol. Auto-pause music during dictation. Music-reactive lighting. Custom dictionary with auto-learn from corrections. Voice-driven gameplay. Eye-tracking integration. Cross-platform support if it makes sense.
Every roadmap item is filtered through the same question: does this reduce a step somewhere?
Free to individuals. Source-available.
Windows only for now. Beta testers welcome — bug reports are the only currency this project accepts.
Free of cost
No in-app purchases. No subscriptions. No usage limits. Download anonymously, dictate a 300-page novel, have Jarvis read it back the next day, repeat as often as you like. There's no fee, no quota, and no point where you start owing money.
Free of surveillance
No account required. No email, no name, no telemetry. Your voice stays on your machine. I can't see what you say even if I wanted to.
Free of lock-in
Source-available under Business Source License 1.1 today. Converts automatically to MIT on January 1, 2030 — meaning if I disappear, get sick, or stop maintaining Samsara, anyone can pick it up and continue. You can't be stranded with software that died with its developer.
The download is only 137 MB zipped. No installer — just extract and run. The CUDA pack is optional — Samsara works fine on CPU only, no NVIDIA GPU needed.
First run
Extract the archive. Run Samsara.exe. A short setup wizard walks you through basic settings on first launch. After that, the tray icon appears. Default wake word is "Jarvis" — say "Jarvis, open Chrome" to test. Full guide and troubleshooting on GitHub.
Licensed under the Business Source License 1.1 — free for personal and non-commercial use. Automatically converts to MIT on January 1, 2030.
Common questions.
Does it work with a quiet voice?
Reasonably well. The wake-word matcher tolerates a wide range. Quiet voices work better with a close-positioned mic — headset, lavalier, or a USB mic within arm's reach.
What about accents?
Whisper handles most English accents better than older speech engines. If a specific word is consistently misheard, Samsara has a custom dictionary you can add to.
Can I use it lying down?
Yes. Wake-word activation means you don't need to reach for a hotkey. A wireless headset or close-positioned mic helps.
Will it conflict with NVDA or other screen readers?
Not intentionally. Samsara doesn't manipulate focus or interfere with screen reader output, but specific edge cases haven't been tested yet. If something breaks, please report it.
Do I need an NVIDIA GPU?
No. CPU works fine. NVIDIA GPUs are faster for transcription if you have one, but not required.
What happens if you stop maintaining Samsara?
The license converts automatically to MIT on January 1, 2030. Even before that, the source is already public. You won't be stranded.
Bugs, ideas, or just hello.
Samsara is early-stage. Things will break. Telling me when they do is the most useful thing you can do for this project.
Samsara is free and stays free. If it saves your hands and you want to send a coffee, I won't say no — but please don't feel you have to.