vipul-modelupload/vani-hinglish-coreml

On-device speech models for Vani (Core ML)

Speech-to-text models converted to Core ML so they run on iPhone, packed as Apple Archives (.aar) for download by the Vani app.

File	Model	Runs with	Size	Licence
hinglish-fast.aar	Oriserve/Whisper-Hindi2Hinglish-Swift (72.6M params, fp16)	WhisperKit	128 MB	Apache-2.0
hinglish-best.aar	Oriserve/Whisper-Hindi2Hinglish-Apex (0.8B params, 4-bit palettized)	WhisperKit	407 MB	Apache-2.0
parakeet-en.aar	nvidia/parakeet-tdt-0.6b-v2, Core ML conversion by FluidInference, unchanged	FluidAudio	440 MB	CC-BY-4.0
parakeet-vocab.aar	NVIDIA Parakeet CTC 110M, Core ML conversion by FluidInference, unchanged	FluidAudio (vocabulary boosting)	86 MB	CC-BY-4.0

Unpack on a Mac with aa extract -i <file>.aar -d <folder>.

Hinglish models

Oriserve's Hindi2Hinglish Whisper fine-tunes, converted with Argmax's whisperkittools. They transcribe Hindi-English speech in English letters ("kal ki meeting mein roadmap discuss karenge"). Each archive is one WhisperKit model folder (AudioEncoder.mlmodelc, TextDecoder.mlmodelc, MelSpectrogram.mlmodelc, tokenizer files). Decode with task transcribe and language en (Oriserve's recipe). Token timestamps are not included: the fine-tunes have no alignment heads.

Parakeet (English)

NVIDIA's Parakeet TDT 0.6B v2, English only, as converted to Core ML by FluidInference. The archive holds exactly the five files FluidAudio loads for AsrModelVersion.v2 (Preprocessor.mlmodelc, Encoder.mlmodelc, Decoder.mlmodelc, JointDecision.mlmodelc, parakeet_vocab.json), unpacked into a folder named parakeet-tdt-0.6b-v2 (FluidAudio's cache name).

Dictionary boost (English)

parakeet-vocab.aar holds NVIDIA's Parakeet CTC 110M as converted by FluidInference (MelSpectrogram.mlmodelc, AudioEncoder.mlmodelc, vocab.json, tokenizer files). FluidAudio's vocabulary boosting uses it to listen for the user's own words and swap them into Parakeet's text where the audio supports them. Unpack into parakeet-ctc-110m-coreml, where FluidAudio looks.

Credit and licences

• Hinglish models: all credit to Oriserve; only the format was changed. Apache-2.0, as the originals.
• Parakeet (both files): models by NVIDIA, Core ML conversions by FluidInference, redistributed unchanged under CC-BY-4.0.