Breeze-ASR-26 Model for CTranslate2

This repository contains the MediaTek-Research/Breeze-ASR-26 model converted to the CTranslate2 format.

The model can be used with CTranslate2 or CTranslate2-based projects such as faster-whisper.

About Breeze-ASR-26

Breeze-ASR-26 (BreezeASR-Taigi) is a Taiwanese Hokkien (Taigi / 台語) automatic speech recognition model fine-tuned from openai/whisper-large-v2. It was trained on approximately 10,000 hours of synthetic Taiwanese Hokkien speech data and outputs Mandarin Chinese character transcriptions.

For more details about the original model, please refer to its model card.

Example

from faster_whisper import WhisperModel

model = WhisperModel("path/to/faster-whisper-Breeze-ASR-26")

segments, info = model.transcribe("audio.wav")
for segment in segments:
    print("[%.2fs -> %.2fs] %s" % (segment.start, segment.end, segment.text))

Conversion Details

The original model was converted with the following command:

ct2-transformers-converter \
  --model /path/to/Breeze-ASR-26 \
  --output_dir faster-whisper-Breeze-ASR-26 \
  --copy_files preprocessor_config.json \
  --quantization float16

Note: The model weights are saved in FP16 format. You can change the type when loading the model using the compute_type option in CTranslate2.

Benchmark Results

The original Breeze-ASR-26 model achieves an average Character Error Rate (CER) of 30.13% on the Taigi ASR Benchmark, outperforming several baseline systems:

More Information

For more information about the original model, please refer to its model card