whisper.cpp

The C/C++ port of Whisper (from Georgi Gerganov, of llama.cpp/ggml) — a plain inference implementation of OpenAI’s model, not a new model. It reimplements Whisper in the ggml tensor library with integer quantization and hardware backends, so the same weights run fast on ordinary CPUs and consumer GPUs with a small memory footprint and no Python runtime.

Why it matters here

It’s the reason local, offline Whisper is practical on a personal machine — the difference between “self-hostable in principle” and “runs on your laptop while you dictate.” That makes it the default backend for consumer-facing tools like vocalinux, which leans on whisper.cpp’s Vulkan GPU acceleration to stay responsive. It’s the inference-efficiency counterpart to whisper‘s model/ecosystem story: whisper explains why the model is the default; whisper.cpp explains how it actually runs on-device. The distinction mirrors quantization in the text-LLM world (cross-wiki) — the same “shrink it to run it locally” move applied to ASR.

Now a baseline others measure against (2026-07)

Microsoft’s VibeVoice-ASR-BitNet claims 1.86× whisper.cpp’s speed at 3 CPU threads, and ships as a llama.cpp-served artifact — the same ggml delivery channel, aimed at the same CPU-only niche. whisper.cpp has become the reference point for “fast enough on a CPU,” which is what a default looks like when the competition starts quoting it. The comparison is vendor-reported and gives no Whisper model size for the baseline, so treat the multiple as a claim, not a measurement.

Related

whisper · vocalinux · speech-to-text · open-source-stt-models · quantization · vibevoice-asr-bitnet