A real Windows desktop app — 13 voices, 6-second voice cloning, AI enhancement, and a batch queue. Runs 100% offline: no cloud, no account, no character caps. Fast mode is free forever; Pro is $12/mo or $89 once.
Full comparison vs ElevenLabs, Murf & PlayHT — limits, cloning, 3-year cost →
We're Cookie Studios — a small independent team building desktop software that respects your computer and your wallet.
VoxWild started in early 2026 because the decent AI text-to-speech options all wanted $22+ a month to rent voices that run fine on a laptop. Nobody had shipped a proper desktop wrapper around Kokoro and Chatterbox — two genuinely great open-source speech models. So we built one. Then we kept building, because every time we used it we noticed something else it needed.
We're small enough that support goes straight to a person who knows the code. Email us if something breaks or you want a feature — we reply within a day, usually within an hour on weekdays.
Yes, but we get why you're asking. SmartScreen warns because code-signing certificates cost ~$400/year and VoxWild is a small indie shop. The installer is hosted on GitHub Releases (public repo, every commit is visible) and the download link on this page points there directly. You can inspect every line of Python before trusting it. We plan to get the exe signed once revenue justifies the cost.
Yes. The only network requests the app ever makes are: (1) a single HTTPS call to Gumroad when you enter a license key, and (2) a daily check to GitHub for new versions. Your text, your audio, your voice clones — none of it leaves your computer. Ever. No account. No cloud.
Yes, on every tier including Free, for both engines. Two caveats: (a) Chatterbox output contains an inaudible neural watermark (a Perth watermark, required by the upstream license — doesn't affect audio quality). (b) The EULA requires that you only clone voices you have rights to clone. Beyond that, what you make is yours.
Honest answer: good enough to fool a listener in a podcast context, not good enough to fool the cloned person's spouse. Longer reference samples (30–60 seconds of clean audio) produce noticeably better clones. The quality of the source audio matters more than the length — a clean 8s sample beats a noisy 60s one.
All twelve questions — refunds, licenses, Mac plans, roadmap →
Drag across the box. Release to hear it. This has nothing to do with the app.