za-anv-multilingual-coreml

ZA ANV multilingual (WhisperKit / Core ML)

Core ML / WhisperKit export of xbiglylabs/za-anv-multilingual-whisper-v3-turbo (fine-tune of openai/whisper-large-v3-turbo on the ANV / Swivuriso data).

Apple Silicon / WhisperKit only — not Transformers, CTranslate2, or faster-whisper.

Uncompressed folder under za-anv-multilingual-v1/ (~1.5 GB). TextDecoderContextPrefill.mlmodelc is not included (optional; conversion PSNR failed on this fine-tune — first tokens may be slower without it).

Converted with argmaxinc/whisperkittools. See NOTICE for upstream credits.

Intended use

On-device ASR for Southern African languages covered by the source model: Tswana (tsn), Zulu (zul), Xhosa (xho), Southern Sotho (sot), Tsonga (tso), Southern Ndebele (nbl), Venda (ven).

Download / load (WhisperKit)

let folder = try await WhisperKit.download(
    variant: "za-anv-multilingual-v1",
    from: "FTruter/za-anv-multilingual-coreml"
)
let pipe = try await WhisperKit(modelFolder: folder.path, download: false)

Or:

hf download FTruter/za-anv-multilingual-coreml --local-dir ./za-anv-multilingual-coreml

let pipe = try await WhisperKit(
    modelFolder: "./za-anv-multilingual-coreml/za-anv-multilingual-v1",
    download: false
)

Force the target language code when you know it; do not rely on autodetection for best results.

Size / device notes

~1.5 GB uncompressed. Low-RAM devices can jetsam — test on target hardware.

Source metrics (from upstream card; not re-measured here)

On the upstream eval set, xbiglylabs report WER 0.1501 / CER 0.0510. Those figures are for the Transformers checkpoint, not this Core ML export.

Limitations

• WhisperKit / Core ML on Apple Silicon only
• Prefill decoder omitted
• Same hallucination / domain limits as the upstream Whisper fine-tune
• This private Hub drop may later move under another namespace

Contents (za-anv-multilingual-v1/)

• MelSpectrogram.mlmodelc, AudioEncoder.mlmodelc, TextDecoder.mlmodelc
• config.json, generation_config.json, preprocessor_config.json
• tokenizer.json, tokenizer_config.json, manifest.json

Root: this card, LICENSE, LICENSE-whisper-apache-2.0.txt, NOTICE.

Citation / credit

Upstream: xbiglylabs/za-anv-multilingual-whisper-v3-turbo. Base: OpenAI Whisper large-v3-turbo (Apache-2.0). Training data: ANV / Swivuriso (dsfsi-anv). See NOTICE.