goodpixelltd/whisper-tiny-cantonese-coreml

Whisper Tiny Cantonese - CoreML

CoreML conversion of Oblivion208/whisper-tiny-cantonese for use with WhisperKit on Apple devices.

Model Description

This is a CoreML-converted and quantized version of the Whisper Tiny model fine-tuned for Cantonese (粵語) speech recognition. The model outputs colloquial Cantonese (口語) rather than written Chinese (書面語).

• Base Model: openai/whisper-tiny
• Fine-tuned Model: Oblivion208/whisper-tiny-cantonese
• Conversion Tool: argmaxinc/whisperkittools
• Format: CoreML (.mlmodelc)
• Quantization: 8-bit palettization (~59MB total)

Performance

• Character Error Rate (CER): 11.10% on MDCC test set (original model)
• Optimized for: Apple Neural Engine (ANE)

License

This model is distributed under the Apache 2.0 License.

Attribution

This work is based on:

1. OpenAI Whisper - Original Whisper model

   • License: MIT License
   • Copyright (c) 2022 OpenAI
   • https://github.com/openai/whisper

2. Oblivion208/whisper-tiny-cantonese - Fine-tuned Cantonese model

   • License: Apache 2.0
   • https://huggingface.co/Oblivion208/whisper-tiny-cantonese
   • Trained on: mozilla-foundation/common_voice_11_0

3. WhisperKit - CoreML conversion tools

   • License: MIT License
   • Copyright (c) 2024 Argmax, Inc.
   • https://github.com/argmaxinc/WhisperKit

Changes Made

• Converted from PyTorch to CoreML format using whisperkittools
• Applied 8-bit mixed-precision quantization for reduced model size
• Packaged with tokenizer and configuration files for WhisperKit compatibility

Usage

This model is designed for use with WhisperKit on iOS/macOS:

import WhisperKit

let whisperKit = try await WhisperKit(
    modelFolder: "path/to/whisper-tiny-cantonese",
    computeOptions: .init(audioEncoderCompute: .cpuAndNE)
)

let result = try await whisperKit.transcribe(audioPath: "audio.wav")

Disclaimer

This model is provided "as is" without warranty of any kind. Performance may vary based on audio quality, accent, and speaking style.