Back to glossary

Term

whisper.cpp

whisper.cpp is an open-source C++ port of OpenAI's Whisper speech recognition — a compact native binary for local transcription on CPU, CUDA, Metal or Vulkan, with no Python dependency.

whisper.cpp — explained in more detail

whisper.cpp is the modern C++ inference implementation of OpenAI’s Whisper model, started in 2022 by Georgi Gerganov — the same developer who launched the parallel local-LLM movement with llama.cpp. Both projects share the ggml tensor library as their backend. Licensed under MIT, repository: ggerganov/whisper.cpp. The library has no Python dependency and compiles to a lean binary that runs on almost any system.

Example / In practice

whisper.cpp ships with several executables: main for file transcription, server as an HTTP endpoint for inference requests, stream for live streaming from a microphone. Models come in the ggml-<size>.bin format (with quantised variants such as q5_0, q5_1, q8_0) and are fetched via the included download-ggml-model.sh script. Supported backends include CPU with AVX/AVX2/AVX-512, CUDA, Metal on Apple Silicon, Vulkan, OpenCL and CoreML through a separate conversion step. Typical use cases are dictation tools, voice input in CLIs, offline subtitling and inference on edge devices without cloud access.

Whisper refers to the OpenAI model itself; whisper.cpp is only one of several runtimes for it — alongside the official Python reference implementation and faster wrappers like faster-whisper (CTranslate2). Trade-offs: the first install compiles locally and needs a compiler toolchain; models have to be picked and pulled explicitly, with no automatic updates.

Entdecke mehr

Themenuebersicht