gzwhisper-kazakh-coreml

Kazakh Whisper Large V3 Turbo — WhisperKit Core ML 4-bit

This repository contains a compiled Core ML conversion for local Kazakh speech recognition on Apple devices. It is used by the open-source GZWhisper iOS application.

Source model

The package is derived from shyngys879/kazakh-whisper-large-v3-turbo, an Apache-2.0 Kazakh fine-tune based on openai/whisper-large-v3-turbo. Please read the source model card for its training description, evaluation, intended uses, and limitations.

Conversion

• WhisperKit Tools 0.4.2
• Forced 4-bit weight quantization
• Outlier decomposition enabled
• Compiled AudioEncoder, TextDecoder, and MelSpectrogram Core ML models
• Tokenizer and configuration included for offline use

Converter correctness checks passed for the decoder (PSNR 38.2) and encoder (PSNR 52.7). These checks measure conversion consistency, not recognition quality. The optional context-prefill converter test was omitted because the source tokenizer metadata was incompatible with that optional test; normal file transcription does not require context-prefill.

Usage

The directory follows the model layout expected by WhisperKit. GZWhisper downloads the folder into its private cache and passes it as an explicit model folder. For Kazakh speech, the application selects the source language code kk and uses Whisper's transcription task.

The converted package is approximately 609 MiB on disk. Peak memory, thermal behavior, real-time factor, and word error rate must be evaluated on each target device and recording domain.

Limitations

Quality may degrade with noise, overlapping speakers, regional accents, code-switching, distant microphones, and long unsegmented recordings. This is an experimental model and must not be represented as universally more accurate than the source or generic Whisper models.

License and attribution

The model package is distributed under Apache License 2.0. LICENSE contains the full license text and NOTICE records the source and conversion changes.