Whisper.cpp
High-performance C/C++ port of Whisper for CPU-based speech recognition.
About
whisper.cpp ports OpenAI's Whisper speech recognition model to plain C/C++ with no external dependencies, a project from Georgi Gerganov built on the same ggml tensor library that later powered llama.cpp. It transcribes and translates audio in the 99 languages Whisper supports while running comfortably on hardware the original PyTorch implementation cannot: phones, Raspberry Pis, and browsers via WebAssembly, alongside macOS, Linux, Windows, FreeBSD, iOS, and Android. Performance comes from mixed F16/F32 precision, integer quantization formats such as Q5_0, zero memory allocations at runtime, and a long list of acceleration paths including ARM NEON, Metal, and Core ML on Apple silicon, AVX on x86, and CUDA, ROCm, Vulkan, and OpenVINO elsewhere. The tooling adds word-level timestamps, voice activity detection, and even karaoke-style subtitle video output. MIT licensed and widely embedded, it is the standard route to shipping offline Whisper transcription inside desktop and mobile applications where a Python runtime is impractical.
Should you use Whisper.cpp?
Pick it when
Pick Whisper.cpp when you embed offline Whisper transcription in a desktop, mobile, or browser app without a Python runtime, or run it on CPUs, Apple silicon, Raspberry Pis, and mixed GPUs with quantized models.
Look elsewhere when
Skip it for GPU batch jobs in a Python stack, where faster-whisper or Insanely Fast Whisper fit more naturally, or when you need speaker labels, which WhisperX adds. Swift apps on Apple devices can use WhisperKit.
Alternatives to Whisper.cpp
- faster-whisper
Python library with CTranslate2 and 8-bit quantization, easier for server pipelines, but it needs the Python runtime whisper.cpp avoids.
- WhisperKit
Swift package on Core ML and the Neural Engine with streaming and a local OpenAI-compatible server, but only on Apple platforms.
- Moonshine
Smaller streaming models for boards and wearables with lower latency than Whisper, but eight languages instead of about 99.
- Vosk
Even lighter 50 MB models with grammars and many bindings, trading Whisper's accuracy for footprint.
Reviews (0)
Leave a Review
No reviews yet. Be the first to review!
Details
- Price
- Free
- Platform
- Local/Desktop
- Difficulty
- Easy (2/5)
- License
- MIT
- Added
- Apr 3, 2026