pocket-tts-cloning-en

Pocket TTS, English voice-cloning weights (non-gated mirror)

Byte-identical mirror of the English voice-cloning weights from kyutai/pocket-tts, redistributed under CC-BY-4.0 so that Yap can ship voice cloning without asking anyone for a Hugging Face account.

Original model © Kyutai Labs, licensed CC-BY-4.0. This mirror adds nothing, removes nothing, retrains nothing and quantizes nothing. If you want the canonical source, get it from Kyutai.

Check it yourself against upstream:

shasum -a 256 languages/english/model.safetensors

---

What this powers: Yap

Talk to your Mac. It talks back. You both yap.

Yap is an open-source macOS menu-bar app that gives your Mac, and the AI agents you work with, two things it was missing:

• Ears. Hold a key and talk. Local Parakeet STT drops your words straight where the cursor is, in any app, running on the Apple Neural Engine.
• A voice. Highlight anything, tap a shortcut, hear it read back in a real Kokoro or Pocket TTS voice. 54 instant voices across 8 languages, or 26 markedly more natural Pocket voices at ~10x realtime on CPU.
• Your voice. Clone your own from a ~20 second clip, then have your Mac read to you in it. That is what these weights are for.

No cloud. No account. No telemetry. No API key. Everything runs on the machine, which is the entire point of the project. You think faster than you type and way faster than you read, so Yap lets you stop doing both.

brew install --cask latent-variable/tap/yap

Source, releases and a 31-second demo: https://github.com/latent-variable/Yap

Why this mirror exists

Yap's promise is that it needs no account. Kyutai gates the cloning weights upstream, which meant the one feature that broke that promise was voice cloning: a Hugging Face login, a terms click, a read token and a Keychain entry, all to fetch a single 209 MB file. It was fragile as well as awkward. A macOS update once evicted the Hugging Face cache and the stored token together, and every cloned voice silently stopped working.

CC-BY-4.0 permits redistribution with attribution, so Yap fetches this mirror like any other model file and verifies the checksum above before using it. No account, nothing to accept, nothing stored.

Acceptable use, and these terms travel with the weights

Kyutai gated the upstream repo to make each user acknowledge this, and redistribution under CC-BY-4.0 does not retire it. It applies here in full, and Yap surfaces it in its own cloning UI:

• Do not clone anyone's voice without their explicit, lawful consent.
• Do not use this for impersonation, fraud or misinformation.
• Do not generate unlawful, harmful or privacy-invading content.

Clone your own voice, or one you have clear permission to use. Nothing else.

Scope

Only the cloning weights are mirrored here. The tokenizer and the non-cloning (catalog voice) weights are not, because they already ship ungated in kyutai/pocket-tts-without-voice-cloning.

Credits

• Model: Kyutai Labs, pocket-tts (paper)
• Mirror + app: Yap by latent-variable