fluister-turbo-coreml

Fluister-turbo (WhisperKit / Core ML)

Same merged whisper-large-v3-turbo weights as digiphyte/fluister-turbo (CTranslate2) and digiphyte/fluister-turbo-transformers (fp16 safetensors), exported for Apple Silicon / WhisperKit (ANE / Metal).

Not a Transformers, CTranslate2, or faster-whisper checkpoint.

Uncompressed folder under fluister-turbo-v2/ (~1.6 GB). TextDecoderContextPrefill.mlmodelc is not included (optional; conversion PSNR sat just under the stock check — first tokens may be slower without it).

Converted with argmaxinc/whisperkittools. Credit DigiPhyte; see NOTICE.

Intended use

On-device speech recognition for Afrikaans and South African English via WhisperKit.

Demo

iOS demonstration app: ftruter/FluisterDemo.

Download / load (WhisperKit)

Model files live in the fluister-turbo-v2/ subdirectory so Argmax-style download works:

let folder = try await WhisperKit.download(
    variant: "fluister-turbo-v2",
    from: "FTruter/fluister-turbo-coreml"
)
let pipe = try await WhisperKit(modelFolder: folder.path, download: false)

Or snapshot the repo and point at the nested folder:

hf download FTruter/fluister-turbo-coreml --local-dir ./fluister-turbo-coreml

let pipe = try await WhisperKit(
    modelFolder: "./fluister-turbo-coreml/fluister-turbo-v2",
    download: false
)

Force language to "af" or "en" for best results; leave autodetection off for this model.

Size / device notes

~1.6 GB uncompressed on disk. Devices with ~4 GB RAM (e.g. some iPhones) can jetsam under memory pressure — test on target hardware.

Contents (fluister-turbo-v2/)

• 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.

Limitations

• WhisperKit / Core ML on Apple Silicon only
• Prefill decoder omitted (see above)
• Sibling ownership under digiphyte/ (next to the CT2 build) may follow separately

Citation / credit

DigiPhyte (Pty) Ltd. See NOTICE and license files. Base: openai/whisper-large-v3-turbo.