Skip to main content

Voice

Mnemosyne OS doesn't just write — she listens and speaks. Turn the sound on and watch a real conversation (asked in the founder's French-accented English, answered aloud):

Voice mode, live: the aurora listens, transcribes as you speak, and answers out loud — transcript on screen.

Voice mode​

Click the Infinity logo (∞) at the top of the app and the aurora takes the screen. From there:

  • Speak naturally — your words are transcribed live below the orb.
  • Say « open notes » — or just chat: commands and conversation mix freely.
  • Say « goodbye » (or press Esc) to end.

The orb answers aloud and in writing — the full reply stays readable on screen while she speaks it.

The voices​

Three output engines, chosen during onboarding or anytime in Settings:

  • System (offline) — your OS's built-in voices, zero download.
  • Local neural — natural voices synthesized in real time, at zero cost, on your own GPU. The engine downloads once, then works offline forever.
  • ElevenLabs (cloud) — bring your ElevenLabs account if you want their voices.
Real time, zero cost — CUDA required

The local neural engine is the sweet spot: real-time synthesis, no API, no bill, nothing leaves your machine — and it needs an NVIDIA GPU (CUDA) to run at that speed. No CUDA? The System voices still work fully offline, and ElevenLabs covers the cloud route.

A Test voice button lets you hear before you commit. And in any chat, the Companion voice toggle (in the options panel) makes replies read themselves aloud.

Clone a voice​

Give Mnemosyne a voice that means something to you: your idol's, a family member's, a departed loved one's — Mnemosyne's voice becomes theirs, for you, on your machine.

  • Five seconds are enough. Any clear recording of the person speaking — a voice message, a video — is enough to start.
  • For the best result, have the voice read the passage below: a quiet room, the best microphone you have, a natural pace. It takes about thirty seconds and covers what the engine needs — questions, numbers, soft and sharp sounds:

Memory is a strange house: some doors open with a sound, others with a scent. This morning I counted three, thirteen, thirty-one steps under a July sky. Would you remember the blue boat, the quiet harbor, the thunder far away? Ask me again tomorrow — I'll keep every detail: the laughter, the whisper, the exact shade of evening. Zero things forgotten. Everything, precisely where you left it.

Your voice stays yours​

The voice engines can also work with a cloned voice — and that power is treated as what it is: the voice is the first sensitive capability in the OS. No app or agent gets access to a voice without an explicit dialog asking you — there is no silent path to it, not even for built-in cartridges.

Where voice shows up elsewhere​

  • MnemoReader reads your PDFs aloud with the same engines.
  • The chat can listen (microphone) and speak (companion voice) without entering full voice mode.