whisper.cpp

// Overview

whisper.cpp is a port of OpenAI's Whisper automatic speech recognition model written in plain C and C++ with no external dependencies. The whole high-level model lives in two files, `include/whisper.h` and `src/whisper.cpp`; everything underneath is the ggml tensor library, the same foundation as llama.cpp. That compactness is the point — it makes the model something you can drop into an application rather than a Python service you have to operate next to one.

The build targets a wide spread of hardware rather than one accelerator. Apple Silicon is a first-class target through ARM NEON, the Accelerate framework, Metal and Core ML, where running the encoder on the Neural Engine is reported to be more than three times faster than CPU-only. Beyond that there are AVX intrinsics for x86, VSX for POWER, and backends for NVIDIA CUDA, AMD ROCm, AMD Ryzen AI NPUs, Vulkan, OpenVINO, Ascend NPUs and Moore Threads GPUs. Inference runs in mixed F16/F32 with no memory allocations at runtime, and integer quantization shrinks the models further.

Because it is just a library and a CLI, it ships everywhere the model needs to go: macOS, iOS, Android, Linux, FreeBSD, Windows, WebAssembly, Raspberry Pi and a Docker image, with bindings including Java and npm. Models are used in the project's own `ggml` format, downloaded by a script in the repo, and voice activity detection is built in. The project is MIT licensed and maintained under the ggml organisation.

// What it does

• Plain C/C++ implementation with zero dependencies and no runtime memory allocation
• Backends for Metal/Core ML, CUDA, ROCm, Vulkan, OpenVINO, Ascend and AMD Ryzen AI NPUs, plus AVX and VSX CPU paths
• Integer quantization (e.g. Q5_0) to cut model size and memory
• Built-in voice activity detection (VAD)
• Runs on macOS, iOS, Android, Linux, Windows, Raspberry Pi, WebAssembly and Docker
• A C-style API in `whisper.h` plus bindings, so it embeds directly in applications

// Getting started

Clone the repo, fetch a model in ggml format, build the CLI with CMake and transcribe a 16-bit WAV file.

git clone https://github.com/ggml-org/whisper.cpp.git
cd whisper.cpp
sh ./models/download-ggml-model.sh base.en

cmake -B build
cmake --build build -j --config Release

./build/bin/whisper-cli -f samples/jfk.wav

ffmpeg -i input.mp3 -ar 16000 -ac 1 -c:a pcm_s16le output.wav

./build/bin/quantize models/ggml-base.en.bin models/ggml-base.en-q5_0.bin q5_0
./build/bin/whisper-cli -m models/ggml-base.en-q5_0.bin ./samples/gb0.wav

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

// When to use it

• Transcribe audio entirely on-device, so recordings never leave the machine
• Embed speech recognition in a desktop, mobile or embedded app through a C API instead of shipping a Python runtime
• Run Whisper on hardware without a CUDA GPU — Apple Silicon, a Raspberry Pi, an NPU or plain CPU
• Build an offline voice command layer or a browser-side transcriber via the WebAssembly build

// How whisper.cpp compares

whisper.cpp alongside other open-source audio, music & voice tools AI/TLDR tracks, ranked by GitHub stars.