saysafe
Copyright 2026 Ratnam Ojha

Licensed under the Apache License, Version 2.0 (see LICENSE).

Third-party models
------------------
saysafe bundles no model weights. With the optional voice extra
(pip install "saysafe[voice]"), these are downloaded or loaded at runtime:

- ECAPA-TDNN speaker embeddings, speechbrain/spkrec-ecapa-voxceleb
  (https://huggingface.co/speechbrain/spkrec-ecapa-voxceleb). Apache-2.0 per its
  model card; trained on VoxCeleb. Downloaded from Hugging Face on first use to
  ~/.cache/saysafe/ecapa.
- Silero VAD (https://github.com/snakers4/silero-vad). MIT. Shipped inside the
  silero-vad package.

The band demo in examples/band_demo (not part of the saysafe package) also uses
faster-whisper with OpenAI Whisper weights (MIT) for speech-to-text, and the Piper
voice en_US-lessac-medium for text-to-speech. That voice is trained on the Blizzard
2013 Lessac dataset, which has its own license: check it before any commercial use.
