leope/ark-asr-3B-mlx

ARK-ASR-3B MLX

Native Apple MLX conversion of AutoArk-AI/ARK-ASR-3B, a multilingual automatic speech recognition model supporting 19 languages.

This repository includes the BF16 checkpoint, tokenizer and feature extractor, native MLX Python inference implementation, conversion code, tests, validation results, and licensing information. PyTorch is not required for inference.

This is a community conversion. It is not affiliated with or endorsed by the original ARK-ASR authors or Apple.

Requirements

• A Mac with Apple silicon
• macOS with Metal support
• Conda
• Approximately 8 GB of available unified memory during inference
• Audio clips no longer than 30 seconds

The runtime currently performs single-clip greedy transcription. This is a custom MLX architecture and is not compatible with mlx-lm, mlx-audio, or Transformers AutoModel.

Installation

hf download leope/ark-asr-3B-mlx --local-dir ark-asr-3B-mlx
cd ark-asr-3B-mlx
conda env create -f environment.yml
conda activate ark-asr-mlx

Transcription

ark-asr-mlx transcribe /path/to/audio.wav --model .

To receive structured output:

ark-asr-mlx transcribe /path/to/audio.wav --model . --json

Python usage:

from ark_asr_mlx import ArkASR

asr = ArkASR.from_pretrained(".")
result = asr.transcribe("/path/to/audio.wav")
print(result.text)

Audio is converted to mono and resampled to 16 kHz. Files longer than 30 seconds are rejected instead of silently truncated.

Native MLX checkpoint

Safetensors is framework-neutral, so the filename alone is not the proof that this is an MLX conversion. This release changes the checkpoint and runs it through a model implemented with native mlx.core and mlx.nn operations:

• all 924 retained tensors load strictly as native mlx.core.array values;
• weights are stored as BF16;
• Conv1d tensors use MLX [out, kernel, in] layout instead of PyTorch [out, in, kernel] layout;
• the tied lm_head.weight and unused learned Whisper position embedding are intentionally omitted;
• inference runs through Metal on Apple silicon without importing PyTorch.

Inspect the checkpoint directly:

python - <<'PY'
import mlx.core as mx

weights = mx.load("model.safetensors")
conv = weights["audio_encoder.whisper.conv1.weight"]
print(type(conv))       # <class 'mlx.core.array'>
print(conv.dtype)       # mlx.core.bfloat16
print(conv.shape)       # (1280, 3, 128), native MLX Conv1d layout
print(len(weights))     # 924
print(mx.metal.is_available())
PY

Conversion provenance

Field	Value
Source	AutoArk-AI/ARK-ASR-3B
Source revision	1e28271b79edc97635783bea65abc89195a09ed3
Format	Native MLX Safetensors
Precision	BF16
Parameter tensors	924
Checkpoint size	7.0 GiB
Output SHA-256	a5e9431bdd648340a40c092e385c4fe1d445d8ad3311dd94609e72af36b0256d

The source shard hashes and complete transform manifest are recorded in conversion.json.

To reproduce the conversion from the pinned upstream revision:

pip install -e '.[dev]'
ark-asr-mlx convert \
  --source AutoArk-AI/ARK-ASR-3B \
  --revision 1e28271b79edc97635783bea65abc89195a09ed3 \
  --output ./converted-model

Run the included unit tests and lint checks with:

pytest
ruff check .

Source parity

The pinned PyTorch source model and native MLX model were evaluated on the same public 12.1-second LibriSpeech sample:

Check	Result
Prompt token IDs	Identical
BF16 input features, max absolute difference	0.0
Adapted audio feature cosine similarity	0.9985675587
Initial decoder logits cosine similarity	0.9999223477
Greedy output token IDs	Identical
Final transcription	Identical

See VALIDATION.md for the transcript and validation command.

Reference benchmark

Measured on an Apple M5 Pro with 24 GB unified memory, macOS 26.6, Python 3.12.13, and MLX 0.32.0:

Measurement	Result
Model load	1.011 s
Audio preprocessing	0.499 s
First token	0.885 s
34-token generation	1.965 s
Generation throughput	17.30 tokens/s
MLX peak memory	7.820 GB

These figures describe one machine and one sample; they are not performance guarantees.

Supported languages

Chinese, English, German, Japanese, French, Korean, Spanish, Polish, Italian, Romanian, Hungarian, Czech, Dutch, Finnish, Croatian, Slovak, Slovene, Estonian, and Lithuanian.

License and attribution

The runtime and upstream model are licensed under Apache-2.0. Please retain the included license, notice, model card attribution, and upstream citation when redistributing this conversion.