shdennlin/breeze-asr-25-ct2

Breeze-ASR-25 — CTranslate2 format for faster-whisper

CTranslate2 (CT2) quantized variants of MediaTek-Research/Breeze-ASR-25, ready to drop into faster-whisper, whisper-streaming, WhisperLiveKit, and any tool that loads CTranslate2 Whisper models.

Breeze-ASR-25 is a Whisper-large-v2 fine-tune by MediaTek Research, optimized for Taiwanese Mandarin and Mandarin–English code-switching (intra- and inter-sentential). On Taiwan-flavored mixed-language input it outperforms vanilla Whisper-large-v2 by a substantial margin while preserving Whisper's English ability.

Available variants

Each subfolder is a self-contained CTranslate2 model directory.

Subfolder	Size	Quant	When to pick
float16/	2.9 GB	fp16	GPU deployment with ample VRAM; baseline quality
int8_float16/ ⭐	1.5 GB	INT8 weights + FP16 compute	Recommended sweet spot — works on both GPU and Mac CPU (see compute_type note below)
int8/	1.5 GB	INT8 throughout	Pure CPU servers; slightly more aggressive quantization

Each subfolder contains the full CT2 layout: model.bin, config.json, vocabulary.json, plus tokenizer + preprocessor files for faster-whisper.

⚠️ Mac CPU compute_type warning

The int8_float16 storage format does not map 1:1 to a runtime compute_type on every device. Mac CPU (Apple Silicon) lacks efficient fp16 compute paths, so you cannot pass compute_type="int8_float16" literally on Mac CPU — you'll get:

ValueError: Requested int8_float16 compute type,
but the target device or backend do not support efficient int8_float16 computation.

Use compute_type="default" and let CT2 pick the best runtime path:

# Pre-download the variant you want (each subfolder is a self-contained CT2 model)
hf download shdennlin/breeze-asr-25-ct2 --include "int8_float16/*" --local-dir ./models

from faster_whisper import WhisperModel

# Recommended — works on both Mac CPU and CUDA GPU
model = WhisperModel(
    "./models/int8_float16",
    device="auto",
    compute_type="default",   # auto: int8_float16 on GPU, int8_float32 on Mac CPU
)

CTranslate2 will print a benign warning on Mac CPU that says it converted the weights to int8_float32 at load time — that's expected and correct behavior.

For NVIDIA GPU you can still pin compute_type="int8_float16" explicitly if you want.

Quick start

faster-whisper

pip install faster-whisper huggingface_hub

Step 1 — download the quantization you want (each subfolder is a self-contained CT2 model):

hf download shdennlin/breeze-asr-25-ct2 --include "int8_float16/*" --local-dir ./models

Step 2 — load and transcribe:

from faster_whisper import WhisperModel

model = WhisperModel(
    "./models/int8_float16",
    device="auto",
    compute_type="default",   # see Mac CPU warning above
)

segments, info = model.transcribe(
    "your-audio.wav",
    beam_size=5,
    language="zh",  # or omit for auto-detect
)
print(f"Detected language: {info.language}")
for seg in segments:
    print(f"[{seg.start:.2f}-{seg.end:.2f}] {seg.text}")

Note: faster-whisper's WhisperModel(repo_id, revision=...) shortcut only works for size aliases (tiny/base/small/...) or Systran-prefixed repos. For custom CT2 repos like this one, pre-download then pass the local path.

Streaming with whisper-streaming

pip install whisper-streaming faster-whisper

from whisper_online import FasterWhisperASR, OnlineASRProcessor

asr = FasterWhisperASR(
    lan="zh",
    modelsize=None,
    cache_dir=None,
    model_dir="./models/int8_float16",   # pre-downloaded CT2 dir
)
online = OnlineASRProcessor(asr)
# ... feed audio chunks ...

Device / compute_type cheatsheet

Device	Recommended subfolder	compute_type	Notes
NVIDIA GPU (Ampere+)	int8_float16	int8_float16 or default	Tensor Cores fully utilized
NVIDIA GPU (older / 8GB VRAM)	int8_float16	default	Lets CT2 pick
Mac CPU (Apple Silicon)	int8_float16	default ⚠️	NEVER literal int8_float16 — fp16 not supported on CPU
x86 CPU (no AVX-512 BF16)	int8	int8 or default	INT8 throughout is most efficient
High-VRAM GPU server	float16	float16	Highest quality

Model details

• Base model: openai/whisper-large-v2 (1.55B parameters)
• Fine-tuned by: MediaTek Research
• Original HF repo: MediaTek-Research/Breeze-ASR-25
• Paper: Breeze ASR 25 / Twister (arXiv 2506.11130)
• Languages: Traditional Chinese (zh-TW), English, Mandarin-English code-switching
• Architecture: Whisper-large-v2 encoder-decoder (32 layers each, n_audio_state=1280)

Conversion provenance

These CT2 variants were produced from MediaTek-Research/Breeze-ASR-25's HuggingFace transformers format using ct2-transformers-converter (CTranslate2 4.7.1), with the standard Whisper tokenizer + preprocessor files copied. Verified on macOS arm64 — faster-whisper successfully transcribes the JFK sample with all three quantization levels.

License

Apache 2.0 — inherited from the upstream Breeze-ASR-25 model and Whisper-large-v2 base. See LICENSE in the original repo.

Companion repo

For whisper.cpp / VoiceInk users (GGML .bin format, Core ML ANE acceleration), see the GGML variants: shdennlin/breeze-asr-25-ggml.

Acknowledgments

Massive thanks to:

• MediaTek Research for releasing Breeze-ASR-25 under Apache 2.0
• OpenNMT / CTranslate2 team for the inference engine
• SYSTRAN for faster-whisper
• OpenAI for the original Whisper model