WhisperKit Hinglish (Core ML, 8-bit linear)
Core ML / WhisperKit pack converted from
Oriserve/Whisper-Hindi2Hinglish-Apex,
then 8-bit linear weight-quantized (AudioEncoder + TextDecoder) for smaller on-device
footprint on Apple Silicon.
Model details
| Property |
Value |
| Base (PyTorch) |
Oriserve/Whisper-Hindi2Hinglish-Apex |
| Architecture |
whisper-large-v3-turbo (~0.8B) |
| Runtime |
WhisperKit / Core ML (ANE-capable) |
| Quantization |
int8 linear weights (linear_symmetric / linear) via coremltools; MelSpectrogram left float |
| Input |
Hindi / Hinglish audio |
| Output |
Romanized Hinglish (Latin script) |
| License |
Apache 2.0 |
Example output: "yaar mai kal office nahi jaunga"
(not "यार मैं कल ऑफिस नहीं जाऊँगा")
Important decoding note
Use language: "en" in WhisperKit decoding options. The fine-tune maps Hindi speech onto
English/Latin tokens; forcing Hindi often breaks romanization.
Contents
Typical layout (folder name may match what you uploaded):
Oriserve_Whisper-Hindi2Hinglish-Apex-w8/
AudioEncoder.mlmodelc
TextDecoder.mlmodelc
MelSpectrogram.mlmodelc
config.json
generation_config.json
Usage — WhisperKit (Swift)
import WhisperKit
let config = WhisperKitConfig(
model: "Oriserve_Whisper-Hindi2Hinglish-Apex-w8",
modelRepo: "YOURUSER/YOUR-REPO" // replace with this repo id
)
let whisperKit = try await WhisperKit(config)
// Prefer language "en" for romanized Hinglish
let options = DecodingOptions(language: "en")
let result = try await whisperKit.transcribe(
audioPath: "audio.wav",
decodeOptions: options
)
print(result.text)
Usage — whisperkit-cli
brew install whisperkit-cli
whisperkit-cli transcribe \
--model-path /path/to/Oriserve_Whisper-Hindi2Hinglish-Apex-w8 \
--audio-path /path/to/audio.wav \
--language en \
--verbose
Or after Hub download, point --model / repo flags at this repository per current whisperkit-cli help.
Quality expectations
- Goal: same features as the FP WhisperKit Hinglish pack (romanized Latin,
language=en).
- 8-bit linear weight quantization is lossy storage; compute remains floating point.
- Expect near-parity with the float Core ML build, not bit-identical transcripts.
- Always spot-check on your own Hindi/Hinglish clips before production use.
How this was built
- Convert
Oriserve/Whisper-Hindi2Hinglish-Apex → WhisperKit Core ML with
whisperkittools.
- Apply
coremltools.optimize.coreml.linear_quantize_weights (int8) to AudioEncoder and TextDecoder.
- Compile with
xcrun coremlcompiler to .mlmodelc.
Credits
License
Apache License 2.0 — same as the Oriserve Apex checkpoint and typical Whisper large-v3-turbo
derivatives. Keep attribution when redistributing.
