whisper.cpp Models for Mobilint NPU

This repository provides all model files needed to run whisper.cpp-mblt, the Mobilint NPU-accelerated fork of whisper.cpp.

Available Files

Model	File	Size	Description
whisper-small	ggml-small.bin	466 MB	GGML model (tokenizer + weights for CPU fallback)
whisper-small	ggml-small-encoder.mxq	93 MB	Mobilint NPU encoder
whisper-small	ggml-small-decoder.mxq	159 MB	Mobilint NPU decoder

Usage

NPU Inference (Mobilint)

# Download all files and run
whisper-cli-mblt \
  -m ggml-small.bin \
  --mxq-encoder ggml-small-encoder.mxq \
  --mxq-decoder ggml-small-decoder.mxq \
  -f audio.wav

# Or auto-download from HuggingFace
whisper-cli-mblt -hf mobilint/whisper-small -f audio.wav

CPU Inference (standard whisper.cpp)

The ggml-small.bin file is also compatible with standard whisper.cpp for CPU-only inference:

whisper-cli -m ggml-small.bin -f audio.wav

Model Details

• Base model: openai/whisper-small (244M parameters)
• Languages: 99 languages supported (English, Chinese, German, Spanish, Russian, Korean, French, Japanese, Portuguese, Turkish, Polish, and more)
• Tasks: Transcription and translation (to English)
• NPU pipeline: Audio → mel spectrogram (CPU) → encoder (NPU, global4) → decoder (NPU, single core, greedy) → text

Related Repositories

• mobilint/whisper-small — Original Mobilint whisper-small model with config.json and tokenizer files
• ggerganov/whisper.cpp — Upstream GGML models for CPU inference

License

Apache 2.0 (same as the original OpenAI Whisper model)