# Sources, licensing and provenance

This is an illustrative public-data ASR reproduction prepared in September 2026. It contains no original Zoice customer recordings or proprietary training data, and does not claim to recover the original model or results.

## Model

- Model: [openai/whisper-tiny](https://huggingface.co/openai/whisper-tiny), multilingual checkpoint.
- Pinned revision: `169d4a4341b33bc18d8881c4b69c2e104e1cc0af`.
- The selected Hugging Face model card labels this artifact Apache-2.0. The original [OpenAI Whisper repository](https://github.com/openai/whisper) is separately distributed under its repository license. Preserve the applicable upstream notices when publishing model artifacts; do not describe the underlying model as proprietary Zoice research.
- Downloaded weights are safetensors; remote model code execution is not enabled.

## Speech data

- Mirror: [dianavdavidson/MUCS-Hinglish](https://huggingface.co/datasets/dianavdavidson/MUCS-Hinglish).
- Pinned revision: `3c66ee127a7b0e0a746d4df1002858567135edc1`.
- Original source: [OpenSLR SLR104 — MUCS 2021 sub-task 2](https://www.openslr.org/104/), Hindi-English technical-lecture speech.
- The mirror card says CC BY 4.0; the original OpenSLR source says **CC BY-SA 4.0**. This package preserves the upstream CC BY-SA attribution and treats redistributed source-derived transcripts/audio under those terms instead of assuming the mirror removed the share-alike condition. It makes no blanket legal claim about the license of resulting model weights.
- Audio and large downloaded shards are excluded from source control by default. Reproduction downloads them from their sources. The selected manifests and prediction files include source-derived reference transcripts and retain source provenance.
- Dataset-derived manifests and prediction text: CC BY-SA 4.0 with attribution to the original MUCS/OpenSLR providers and the Hugging Face repackager. Original scripts in this package are separately licensed as stated in `LICENSE`.

The corpus contains technical tutorials, not call-center conversations. The dataset-provided `speaker_id` and session IDs are used for split checks; the package does not independently identify individual speakers. Near-duplicate audio and overlap with Whisper pretraining are not fully audited.

## Method

The method is informed by [Hugging Face's multilingual Whisper fine-tuning tutorial](https://huggingface.co/blog/fine-tune-whisper) and its [ASR evaluation guide](https://huggingface.co/learn/audio-course/en/chapter5/evaluation). The local pipeline uses its own explicit PyTorch training loop and evaluation records. It does not reproduce the tutorial's published benchmark or claim its results.

Dependencies and their exact installed versions are recorded in `requirements-macos-py312.lock.txt`; their upstream licenses remain applicable.