whisper.cpp
A local Whisper speech recognition inference project based on ggml, created by Georgi Gerganov. It focuses on efficient local inference for the Whisper family of models.
Core Facts
Timeline (last 90 days)
To integrate an inference engine in C++, a common approach is to use backends such as ONNX Runtime, whisper.cpp, or TensorFlow Lite, converting the model to ONNX or GGUF format before loading it.
There are three viable paths for self-hosting Whisper: whisper.cpp (a C/C++ port, with no Python dependency), faster-whisper (based on CTranslate2, several times faster), and the reference implementation openai/whisper (the original PyTorch version).
whisper.cpp takes only a few minutes to compile, can run on CPU, and supports Metal acceleration on Apple Silicon and CUDA acceleration on Nvidia.
With the large-v3 model, transcription speed on a modern laptop CPU using whisper.cpp is about 1–2 times real time, about 3–8 times on Apple Silicon with Metal, and about 15–30 times on a consumer-grade Nvidia GPU using faster-whisper.
Running the Whisper model on iOS devices requires converting the model to Core ML or using inference frameworks such as whisper.cpp, and leveraging the Apple Neural Engine for acceleration.
围绕 Whisper 构建了 faster-whisper、whisper.cpp 等第三方工具生态
whisper.cpp is a C/C++ implementation based on the GGML tensor library that supports pure CPU inference and can achieve real-time transcription on Apple M-series chips.
All Facts (9)
Quantized versions of Whisper (such as whisper.cpp and faster-whisper) can run speech recognition in real time on consumer-grade CPUs.
65%Unverified在移动端集成Whisper通常使用其量化压缩版本（如whisper.cpp）以平衡识别精度与设备性能，无需将音频数据上传至云端处理
80%Unverifiedwhisper.cpp is a C/C++ implementation based on the GGML tensor library that supports pure CPU inference and can achieve real-time transcription on Apple M-series chips.
60%UnverifiedTo integrate an inference engine in C++, a common approach is to use backends such as ONNX Runtime, whisper.cpp, or TensorFlow Lite, converting the model to ONNX or GGUF format before loading it.
50%UnverifiedThere are three viable paths for self-hosting Whisper: whisper.cpp (a C/C++ port, with no Python dependency), faster-whisper (based on CTranslate2, several times faster), and the reference implementation openai/whisper (the original PyTorch version).
50%Unverifiedwhisper.cpp takes only a few minutes to compile, can run on CPU, and supports Metal acceleration on Apple Silicon and CUDA acceleration on Nvidia.
50%UnverifiedWith the large-v3 model, transcription speed on a modern laptop CPU using whisper.cpp is about 1–2 times real time, about 3–8 times on Apple Silicon with Metal, and about 15–30 times on a consumer-grade Nvidia GPU using faster-whisper.
50%UnverifiedRunning the Whisper model on iOS devices requires converting the model to Core ML or using inference frameworks such as whisper.cpp, and leveraging the Apple Neural Engine for acceleration.
50%Unverified围绕 Whisper 构建了 faster-whisper、whisper.cpp 等第三方工具生态
50%