kenkui-voices

95 pre-compiled speaker embeddings for pocket-tts, built from kyutai/tts-voices prompts.

• voices.json — the manifest: what ships, what it was compiled from, and the pocket-tts range it works with.
• compiled/ — the embeddings, addressed by the revision in voices.json.
• previews/ — one rendered line per voice.
• tools/ — rebuild, verify, and publish.

Rebuilding

Compiling needs the gated kyutai/pocket-tts weights. Accept the terms and authenticate first.

hf download D1zzl3D0p/kenkui-voices --repo-type dataset --local-dir kenkui-voices
cd kenkui-voices/tools && uv sync
uv run python rebuild.py --all --force    # rebuild everything
uv run python verify.py                   # does this pocket-tts still match?

Rights

Every voice records its own license_id, commercial_use_allowed, and voice_rights. VCTK voices are CC-BY-4.0; EARS voices are CC-BY-NC-4.0. Both stay non-commercial by default: review the corpus terms and speaker consent for your own use.