Community setup: run VoiceStudio fully on a free Colab T4 (UI on your PC, GPU in the cloud)

[NurMohammadDuari]

Disclosure first: this is a third-party community project, not an official
VoiceStudio release. No app code is modified or vendored — it is a notebook, a
launcher and docs that use the remote-backend mode the app already ships.

What it is

VoiceStudio assumes the GPU is in the same machine as the UI. If yours isn't
there — or you just don't want to buy a card — you can run the compute on a
free Google Colab T4 and leave the PC doing nothing but rendering.

VoiceStudio already supports this (docs/remote-gpu.md, Settings → Sharing →
Remote backend). What's missing is the fiddly part: pairing a URL with a key
that keeps working, surviving a Colab recycle without re-downloading 5 GB of
models, knowing which model pack each tier actually wants, and not losing an
afternoon to a 403 that turns out to be a by-design loopback gate.

What it contains

• A 6-box Colab notebook — install, start, one-time setup, reprint your
  credentials, optional dubbing pack
• Key persistence on Google Drive — the API key survives a VM recycle, so the
  copy already saved on your PC stays valid; only the random tunnel URL changes
• Self-verification — box 2 refuses to print a URL/key pair until it has
  proved the exchange works through the public tunnel, using the same handshake
  the desktop app performs
• GPU Turbo — enables OMNIVOICE_FLASHINFER (the app's own opt-in FlashInfer
  path) with a torch-pinned install and an import test, skipped safely when the
  wheel does not exist for your CUDA build
• A launcher that verifies a saved key against the backend and asks for one
  only when it is actually rejected
• Docs including a source-level analysis of the auth gates and route map, so
  it's clear which limitations are security and which are bugs

Honest limits

Colab caps GPU hours and disconnects idle tabs. The default TTS model is CC-BY-NC,
so this is not a path to reselling a dubbing service. A handful of buttons stay
disabled because they act on the backend host's filesystem and can never address
your PC over a tunnel — saving audio is an HTTPS download and is unaffected.

Nothing here is Colab-specific: point the same launcher at any GPU box and it is
the same setup.

👉 https://github.com/NurMohammadDuari/VoiceStudio-Colab-GPU

Feedback welcome — particularly from anyone running it on AMD/ROCm, or on a
rented GPU rather than Colab, since those paths are documented but untested.