Modelpublic

Printman1078/whisper.cpp

sourceHugging Faceupdated 1y agoView on Hugging Face

0likes

Model Card

whisper.cpp Windows Universal Binary Package

This repository provides a community-compiled whisper.exe for Windows x64, requiring no local compilation. It supports a wide range of NVIDIA GPUs (from Maxwell sm50 to Hopper sm100) and automatically falls back to CPU when no GPU is available.

Why Recompile?

The official prebuilt binaries typically include GPU kernels for common hardware (e.g., sm86). Older or less common GPUs (like the GTX 960M with sm50) lack corresponding kernels and cause runtime errors such as:

CUDA error: no kernel image is available for execution on the device

To solve this compatibility gap, we rebuilt with the following CMake options:

bash

-DCMAKE_CUDA_ARCHITECTURES=50^;52^;53^;60^;61^;62^;70^;72^;75^;80^;86^;87^;89^;90^;100^

—CMAKE_CUDA_ARCHITECTURES=… Statically compiles kernels for 19 compute capabilities (from sm50 to sm100), ensuring nearly all NVIDIA GPUs work out of the box.
—GGML_CUDA_DYNARCH=OFF Disables runtime architecture probing and relies solely on the above static list.
—BUILD_SHARED_LIBS=OFF Statically links all components (whisper, ggml, ggml-cpu, ggml-cuda) into a single whisper.exe, eliminating runtime plugin scanning and improving startup speed.

Included Files

After extracting whisper-win.zip, you only need these four files:

File	Description
`whisper.exe`	Main executable (GPU/CPU)
`cudart64_12.dll`	CUDA Runtime 12.x
`cublas64_12.dll`	cuBLAS 12.x
`cublasLt64_12.dll`	cuBLAS Lt 12.x

These DLLs come from the CUDA Toolkit bin/ directory; you do not need a full CUDA installation.

Quick Start

1.Download a model

bash

   curl -L -o model.bin \
     https://huggingface.co/ggerganov/whisper.cpp/resolve/main/models/ggml-base.en.bin

1.Extract and navigate

bat

   unzip whisper-win.zip -d whisper-win
   cd whisper-win

1.Run transcription

bat

   whisper.exe -m model.bin -f audio.wav --language zh -otxt

—-m: path to the model file
—-f: audio file to transcribe (WAV/MP3)
—--language: e.g. zh, en
—-otxt: output as plain text

1.Automatic CPU fallback If no NVIDIA GPU or driver is found, the program seamlessly uses the CPU backend.

System Requirements

—OS: Windows 10 or 11 (x64)
—NVIDIA Driver: Compatible with CUDA 12.x (recommend R560 or newer)
—Visual C++ Redistributable: Microsoft Visual C++ 2015–2022

License

—whisper.cpp & ggml: MIT License (See the original LICENSE)
—CUDA Runtime & cuBLAS: NVIDIA EULA (retain NVIDIA’s license notices when redistributing)

Feel free to open issues if you encounter any problems or have suggestions!