weiren119/Breeze-ASR-25-coreml-int8

Breeze-ASR-25 CoreML INT8

INT8 weight-quantized CoreML version of MediaTek-Research/Breeze-ASR-25 for on-device speech recognition on Apple Silicon.

Model Files

Component	Size	Description
AudioEncoder.mlmodelc	609 MB	Encoder (INT8 quantized)
TextDecoder.mlmodelc	867 MB	Decoder (INT8 quantized)
Total	1.48 GB

Compression Details

• Method: Linear symmetric INT8 weight quantization via coremltools.optimize.coreml.linear_quantize_weights
• Storage precision: Mixed (Float16, Int8)
• Original model size: ~2.9 GB (Float16)
• Compression ratio: ~2x

Model Architecture

• Base: Whisper-large-v2 fine-tuned for Taiwanese Mandarin + English code-switching
• Encoder input: logmel_data (1 x 80 x 3000)
• Encoder output: output (1 x 1500 x 1280)
• Decoder input: token_data (1 x 1) + audio_data (1 x 1500 x 1280)
• Decoder output: logits (1 x 1 x 51865)
• Spec version: 7 (iOS 16+ / macOS 13+)

I/O Naming Convention

Uses whisper.cpp naming convention:

• Encoder: logmel_data -> output
• Decoder: token_data + audio_data -> logits

System Requirements

• macOS 13+ / iOS 16+
• Apple Silicon (M1/M2/M3/M4 or A-series)

Conversion

Converted using the sheep52031/breeze-asr-25-coreml-ane conversion toolchain with a custom HuggingFace decoder wrapper.

python convert-whisper-to-coreml_int8_support.py \
  --model MediaTek-Research/Breeze-ASR-25 \
  --hf-model \
  --quantize-int8

License

Apache 2.0 (same as the original Breeze-ASR-25 model).

Acknowledgments

• MediaTek Research - Breeze-ASR-25 model
• sheep52031 - Conversion toolchain
• Apple coremltools - CoreML conversion and optimization