Skip to content

Voice models and licensing (important)

AbstractVoice (this library) is MIT-licensed, but voice models / weights are licensed separately.

This document is intentionally conservative: it does not attempt to “summarize” every upstream license. For any serious use (especially commercial), verify license terms at the source of the model you download.

Piper voices (default TTS)

AbstractVoice uses Piper for cross-platform, dependency-light TTS.

  • Piper voice weights are sourced from the upstream Piper voice catalog (commonly distributed via rhasspy/piper-voices).
  • Licenses vary per voice (dataset + model card).
  • Cache directory (by default): ~/.piper/models

Recommended verification workflow:

1) Download the voice you intend to ship/use:

python -m abstractvoice download --piper en

2) Record which voice/model files you distribute (filenames under ~/.piper/models). 3) Verify the upstream license/model card for that exact voice.

STT models (faster-whisper)

STT weights are fetched through the Faster-Whisper / HuggingFace caching mechanisms. Licensing is model-dependent (e.g. Whisper variants).

For offline deployments, prefetch explicitly:

python -m abstractvoice download --stt small

Voice cloning (optional)

AbstractVoice supports voice cloning behind optional extras:

  • abstractvoice[cloning] (OpenF5-based; large artifacts)
  • abstractvoice[chroma] (Chroma-4B; very large; GPU-heavy)
  • abstractvoice[audiodit] (LongCat-AudioDiT-1B; large weights via HF)
  • abstractvoice[omnivoice] (OmniVoice; very large; torch/transformers)

Licensing is engine- and model-dependent; verify:

  • the Python package license (code)
  • any model weights / checkpoints you download
  • any dataset-specific restrictions

Note on vendored code:

  • This repo includes a derived implementation of LongCat-AudioDiT under abstractvoice/audiodit/* to avoid trust_remote_code.
  • Upstream license text is included in third_party_licenses/longcat_audiodit_license.txt.

For AudioDiT specifically, verify the model card/source for:

python -m abstractvoice download --audiodit

For OmniVoice specifically, verify the model card/source for:

python -m abstractvoice download --omnivoice

Practical guidance

  • Treat model weights as third-party assets with their own compliance requirements.
  • Pin exact versions/revisions in production and record them for audits.
  • If you distribute models with your product, include the required notices.