whisper-podlodka-turbo-MLX

Whisper-Podlodka-Turbo (MLX)

Apple MLX checkpoints for bond005/whisper-podlodka-turbo (Apache-2.0). PyTorch / Transformers checkpoints remain on the upstream Hub repo; this repo only holds converted weights.safetensors plus config.json per variant.

Conversion followed mlx-examples/whisper/convert.py (python convert.py --torch-name-or-path bond005/whisper-podlodka-turbo ...). Output filenames were renamed from model.safetensors to weights.safetensors so they load with mlx-whisper on PyPI.

---

Setup

pip install mlx-whisper

Audio decoding in mlx-whisper uses ffmpeg for many formats; install it from your OS package manager.

---

Inference

The Hub repo contains three subfolders. Point path_or_hf_repo at the folder that holds weights.safetensors and config.json (download once, then choose fp16, q8, or q4).

from pathlib import Path

from huggingface_hub import snapshot_download
import mlx_whisper

repo = "evilfreelancer/whisper-podlodka-turbo-mlx"
root = Path(snapshot_download(repo))

variant = "q8"  # or "fp16", "q4"
model_dir = root / variant

result = mlx_whisper.transcribe(
    "audio.wav",
    path_or_hf_repo=str(model_dir),
    language="russian",
)
print(result["text"])

CLI (mlx-whisper):

hf download evilfreelancer/whisper-podlodka-turbo-mlx --local-dir ./mlx-podlodka
mlx_whisper audio.wav --model ./mlx-podlodka/q8 -l ru

Adjust --model to ./mlx-podlodka/fp16 or ./mlx-podlodka/q4 as needed.

---

Hardware

mlx-whisper targets Apple Silicon first; MLX also supports Linux with CUDA or CPU wheels. Use the same MLX install flavor as your platform (MLX install docs).

---

Model card (upstream)

The ASR model is Whisper-Podlodka-Turbo, a fine-tuned Whisper large-v3-turbo focused on Russian (plus English), punctuation, and robustness. Full description, metrics, training data, evaluation tables, and citation are on bond005/whisper-podlodka-turbo.

Citation

@misc{whisper-podlodka-turbo,
  author = {Ivan Bondarenko},
  title = {Whisper-Podlodka-Turbo: Enhanced Whisper Model for Russian ASR},
  year = {2025},
  publisher = {Hugging Face},
  journal = {Hugging Face Model Hub},
  howpublished = {\url{https://huggingface.co/bond005/whisper-podlodka-turbo}}
}