Skip to content
nikita prudnikov/RU: Читать по-русски
All projects

Platforms: Windows · Linux · macOS

Not yet public

Hyperscribe Companion

Runs Hyperscribe’s speech recognition on your computer’s GPU, for faster captions.

Availability

This project is not publicly available yet. You can explore the demo and project details, or contact me about it.

Ask about this project
Hyperscribe Companion’s window three times: the welcome screen with the model download, the live transcript, and the compute settings

What it is

An optional desktop accelerator for Hyperscribe. Your phone sends the glasses’ microphone to your computer over Tailscale; Companion recognises the speech on its GPU and sends the captions back to the phone and the lens.

One model on every platform: NVIDIA Nemotron 3.5 ASR, 32 language-locales including Russian, Ukrainian, Chinese, Japanese and Korean, with punctuation. Words reach the phone about a tenth of a second after they are spoken.

Hyperscribe stays usable without it: the phone keeps an offline backup and reconnects when the computer returns. Companion keeps audio and the live transcript in memory only; recordings are saved on the phone.

What it does

  • Any GPU, or none

    NVIDIA through CUDA, AMD through ROCm or Vulkan, Intel through Vulkan, Apple silicon through Metal. With no usable GPU it runs on the CPU, still in real time on a modern machine.

  • Streaming, settled fast

    Words are settled about 0.3 s after they are spoken and are never rewritten after that.

  • Private by default

    It listens only on this computer and on your tailnet through Tailscale’s HTTPS, never on the public internet. Plain local Wi-Fi access is off unless you turn it on.

  • Headless servers too

    On a GPU box without a display, it prints the pairing address and a QR code in the terminal.

Screens

Welcome: This computer does the listening, the speech model and the Download 751 MB and start button
Live: the transcript the phone receives, with the device and the time per refresh
Settings: the model, the compute backend and the server

What you need

  • Windows 10 or 11, Linux (x86_64 or arm64), or a Mac with Apple silicon
  • A GPU is recommended (NVIDIA, AMD, Intel or Apple), but not required
  • About 1 GB of free disk space for the speech model
  • Tailscale on the computer and the phone, and Hyperscribe on your G2

What’s new

  1. Coming next

    • One model everywhere: NVIDIA Nemotron 3.5 ASR, with streaming that settles words in about 0.3 s.
    • Windows and Linux support, with CUDA, ROCm, Vulkan, Metal and CPU.
    • A headless mode for machines without a display, and a self-test.

  2. 0.1.0

    • First source release for Apple-silicon Macs, with QR pairing over Tailscale.

More projects

  • Hypernova

    Manage Claude Code sessions from Android, macOS and Even G2 glasses.

  • Hyperscribe

    Live captions on your G2 glasses. Speech recognition stays on your own devices.

  • Hypernavigate

    Offline 3D walking navigation on G2 glasses. Saint Petersburg first.