# Third-party notices - Align

The bundled Align models were trained from speech and machine-generated alignment references.
The source datasets and reference systems are not redistributed in this package.

## Training audio

- **FLEURS** - Google - **CC BY 4.0**. Multilingual speech used for the production model.
  Dataset: https://huggingface.co/datasets/google/fleurs

## Reference generation

- **Qwen3-ForcedAligner-0.6B** - Alibaba Qwen team - **Apache-2.0**. Primary word-boundary
  reference for all nine languages. The aligner is not included here.
- **OWSM-CTC v4 1B** - ESPnet/WavLab contributors - **CC BY 4.0**. Used on the validation
  split to estimate CTC timing offsets and as a gross alignment-outlier detector where stable.
  OWSM timestamps are not averaged into the final references. The model is not included here.
- **ESPnet** - ESPnet contributors - **Apache-2.0**. CTC inference and alignment tooling used
  by the training pipeline.

## Runtime frameworks

Align links only Apple system frameworks, including Core ML, Accelerate, AVFoundation, and
Speech. No third-party runtime library is bundled.

The compiled Align weights, calibration policy, and Swift implementation are distributed under
[`LICENSE.md`](LICENSE.md). Dataset and reference-model licenses remain with their respective
copyright holders.