piper-voices-GGUF

Piper voices — GGUF bundle

rhasspy/piper VITS voices converted to ggml's GGUF format (arch=piper) for the CrispStrobe/CrispASR native runtime. Each voice is a single self-contained file — the phoneme-id map and the espeak-ng voice are embedded in the GGUF, so no companion is needed. Output is 22.05 kHz mono (CrispASR resamples to its 24 kHz playback convention).

Tiny and fast: ~15–60 MB per voice, single-digit-ms-per-sentence on CPU — the best fit for mobile (CrisperWeaver) and quick previews. Converted with models/convert-piper-to-gguf.py (F16; reads the upstream .onnx + .onnx.json).

Voices

Only permissively-licensed voices are hosted here — the underlying training datasets allow redistribution (CC0 / public domain). Restrictive voices from rhasspy/piper-voices (e.g. en_US-lessac, Blizzard 2013 research license; the CC BY-NC-SA voices) are deliberately excluded.

The Thorsten + kerstin German voices are released CC0 (public-domain dedication — Thorsten-Voice / Thorsten Müller; kerstin from rhasspy/dataset-voice-kerstin). en_GB-cori is public domain (trained on LibriVox recordings, which are public domain).

† CC-BY 4.0 — attribution required. mls is from Multilingual LibriSpeech (MLS); en_US-libritts_r is from LibriTTS-R (Koizumi et al., Google). Credit the dataset per CC-BY 4.0 when shipping their audio.

‡ M-AILABS — BSD-style license. eva_k / karlsson / ramona are from the M-AILABS Speech Dataset (German voices derived from LibriVox + Project Gutenberg public-domain sources). Its BSD-style license permits commercial use and redistribution provided the copyright notice is retained, and forbids using the contributors' names to endorse derived products. Retain the M-AILABS notice when redistributing.

Excluded (license not redistributable)

Deliberately not converted/hosted here:

• en_US-lessac — Blizzard 2013 (CSTR Edinburgh), research/non-commercial.
• en_US-ryan, en_US-hfc_female/male, de_DE-pavoque — CC BY-NC-SA (non-commercial).

Licensing

Two layers, both permissive here:

• Runtime + converter — the Piper architecture, the espeak-ng phonemizer integration, and convert-piper-to-gguf.py are MIT (rhasspy/piper is MIT-licensed; CrispASR's runtime is its own).
• Voice weights — each GGUF is a derivative of an upstream Piper voice and carries that voice's dataset license (the table above). All voices here are redistributable and commercial-use-OK, but obligations differ: CC0 / public-domain (thorsten, kerstin, cori) need nothing; CC-BY 4.0 (mls, libritts_r) need dataset attribution; M-AILABS (eva_k, karlsson, ramona) need the copyright notice retained. Honour the per-voice terms for any audio you ship. Crediting Thorsten-Voice and the Piper project is appreciated.

To add a voice with a different license (e.g. the CC BY 4.0 en_US-libritts_r), convert it yourself and honour that voice's terms — do not assume the repo-level tag applies.

Usage

crispasr --backend piper -m piper-de_DE-thorsten-medium-f16.gguf \
  --tts "Guten Tag, dies ist ein Test." --tts-output out.wav

In CrisperWeaver these appear in the Synthesize screen's model picker once downloaded; the backend resamples 22.05 kHz → 24 kHz transparently.

Voice provenance (EU AI Act Art. 50(4))

Every piper voice is a single-speaker VITS model trained on one named donor's recordings: the Lessac corpus, Thorsten Müller (who released his voice under CC0), Kerstin. Releasing a voice is consent to publish it — it is not a reason to stop telling listeners the audio is synthetic.

CrispASR records this as speaker_identity=real_person. Output synthesized with it carries a spoken AI disclosure, because audio resembling an identifiable person is a deep fake under Art. 3(60) whether or not any cloning took place. It does not require --i-have-rights: the donor's agreement to the training is a licensing matter settled upstream, which a downstream operator cannot attest to.

Override per run with --speaker-identity, or stamp a file permanently with models/stamp-speaker-identity.py. See docs/eu-ai-act.md §6.2a.