A glowing orb of light lives on your screen, hears its name across the room, answers with a real neural voice, runs your whole-home music, remembers your life, and can even teach itself new skills — all on your own hardware. No cloud. No subscriptions to talk to your own house.
The orb tells you exactly what she's doing — no guessing, no blinking confusion.
Vesper isn't a speaker with a bolted-on assistant. She's a companion designed from the wake word up — engineered so the TV never answers, the music never interrupts, and your voice always does.
Always-on listening with a wake word verified by two speech engines. Whisper re-listens to the last second of audio to confirm — she wakes ~0.4 s after you finish saying “Vesper”, and noise or near-words can't trigger her by accident.
Speech recognition runs on your own server — Vosk + Whisper, fully local. Nothing is sent to Google or anyone else. Vesper keeps working with the internet unplugged.
Kokoro-82M neural text-to-speech — 4.18 UTMOS naturalness, benchmarked against every open model. She starts speaking seconds earlier by streaming sentence by sentence, and 54 voices let you pick hers.
Say “Vesper” while she talks and she stops mid-sentence to listen. Random chatter never cuts her off — and her own voice from the speakers never echoes back into the transcript.
An adaptive noise gate learns your room's hum, a tonal-run detector ignores music and TV drone, and one-click calibration pins the threshold to your microphone. Your voice over the music still gets through.
Voiceprint speaker lock: whoever wakes Vesper owns the conversation. Windows of audio that clearly belong to someone else are quieted before transcription — the household argues, Vesper listens to you.
She distills the things worth keeping — your preferences, your dog, your plans — into a durable memory file injected into every turn. Say “remember…” or let her learn quietly in the background. You can read exactly what she knows.
“Vesper, play some jazz in the kitchen.” Full Roon control by voice — zones, volume, skip, pause. While music plays, the chat slides away, a now-playing panel appears, and the orb dances to the beat.
“Play the Bohemian Rhapsody music video” and it appears on your screen — auto-advancing through embeddable results, with the chat still usable. She even checks if you're watching and moves the video out of the chat's way.
Ask “what is this?” and the webcam wakes just long enough to see, then sleeps. Flip the camera off and she never touches it. Your eyes stay yours.
“Vesper, learn a skill that tracks my water” — and she writes the plugin herself: logic, voice hooks and a custom on-screen overlay, in one portable JSON file. Install by voice, export, import, share.
After she answers, a wake-free window stays open — just keep talking. Pauses mid-thought stay one request. Say “goodbye” and she sleeps. Web search, timers, files and facts, woven into natural flow.
Every picture below is a live capture of the running product — not a mockup.
Big-tech assistants ship your voice to data centers. Vesper runs on a machine you own — the browser only ever streams audio to your server, and the 🧠 toggle can move recognition into the browser itself, so a single byte of audio never leaves the device.
Skills are single-file JSON plugins — and Vesper can write them herself. Each skill gets sandboxed server code, a persistent store, a voice, and its own overlay UI on your screen.
Vesper Standalone runs everything locally on your Windows PC — no cloud account, no audio leaving your machine. One installer, one click, and she's alive.
First launch downloads her models (~2–4 GB) and sets up everything else automatically.
You're about to meet Vesper while she's still a work in progress — brilliant, but occasionally endearingly clumsy. You may bump into the odd bug along the way; consider it part of the adventure. Every report you send helps her glow a little brighter, so please tell us what you find.
Vesper launches with one-time pricing — a single payment, no subscriptions, no per-seat fees, no cloud metering. Buy it once and she lives in your house for good.
Home, Household and Studio editions — each a single payment that includes the full voice experience, offline recognition, memory and skills. Exact tiers are being finalized now; backers get the earliest, best pricing.
Kickstarter coming soon →Yes. Wake-word detection, speech recognition and text-to-speech all run on your own machine. Without internet she loses only web search and YouTube videos — conversation, music control, memory and skills keep working.
Any machine from the last decade as the server (it happily runs on CPU), plus any Chrome or Edge browser as a “screen” — an old laptop, a kitchen tablet, a desktop. Microphone required for voice; camera optional for vision.
No. Two-engine wake verification, an adaptive noise gate and a music/tonal detector all stand between your TV and Vesper. Your voice over that same music still wakes her.
Any OpenAI-compatible model with vision and tool-calling — run a local model for total privacy or point her at a hosted one. Swap it with one environment variable.
One listener per machine — the newest tab owns the mic and older tabs politely go quiet with a “take over” button. Different devices (phone + PC) each keep their own session, sharing the same owner voiceprint and memory.
Any time: ⚙️ options → memory → view facts shows the exact file she reads every conversation. Edit it or delete it; it's plain Markdown on your share.
The Kickstarter launches soon — backers get the best one-time price, forever.
Kickstarter coming soon →