whisper.cpp
C/C++ port of OpenAI's Whisper: offline speech-to-text on almost any device
Overview
whisper.cpp is Georgi Gerganov's dependency-free C/C++ port of OpenAI's Whisper speech-recognition model, built on the same ggml library that powers llama.cpp. It transcribes audio fully offline on CPUs, Apple Silicon, GPUs, phones, Raspberry Pis and even in the browser via WebAssembly, using compact quantized ggml model files.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 22 Jul 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
What you can produce with whisper.cpp
- Transcribe a podcast or meeting recording locally with whisper-cli and a quantized large-v3-turbo model, with no audio ever leaving the machine.
- Run continuous real-time microphone transcription with the whisper-stream example, sampling audio every half second on a laptop CPU.
- Stand up a self-hosted HTTP transcription endpoint with whisper-server's OpenAI-style API using the official ghcr.io/ggml-org/whisper.cpp Docker image.
- Ship on-device dictation inside an iOS or Android app via the precompiled XCFramework or the whisper.rn React Native binding.
- Cut Apple Silicon transcription time by about three times by generating a Core ML encoder that runs on the Neural Engine.
- Reduce silence-induced hallucinations and total processing time by enabling the built-in Silero VAD with the --vad flag.
- Produce word-level timestamps with -ml 1 or generate karaoke-style subtitle videos with the -owts output option.
ASEAN Perspective
whisper.cpp in Southeast Asia
ASEAN-region availability and pricing notes coming soon. Drop the editorial team a note via /contact/ if you can supply local context (Singapore/Malaysia/Indonesia/Thailand/Vietnam).
whisper.cpp is for turning speech into text on hardware you control. It compiles to a lightweight dependency-free binary, loads a single ggml model file — tiny (75 MiB) through large-v3, plus q5_0 quantizations — and transcribes fully offline, with Metal, Core ML, CUDA, Vulkan, ROCm and OpenVINO acceleration available per platform. The bundled whisper-server exposes an OpenAI-style HTTP endpoint, and word-level timestamps, translation and native Silero VAD are built in.
With bindings for Rust, Go, Ruby, Java, .NET, React Native and Unity, it's a good fit for developers embedding transcription in their apps, privacy-conscious users transcribing interviews and meetings locally, and edge deployments — phones, Raspberry Pis, air-gapped machines — where a Python stack or a cloud API is a non-starter. That said, faster-whisper remains the stronger pick for anyone with an NVIDIA GPU who needs maximum batch throughput, and anyone needing named speakers must pair it with a separate diarization tool.
Free and MIT-licensed end to end: the code, the converted Whisper models (OpenAI published the weights under MIT) and the official Docker images (CPU, CUDA, Vulkan and MUSA variants). There is no hosted or paid product — build from source with CMake, pull ghcr.io/ggml-org/whisper.cpp:main, install via Conan, or consume it through the official npm WebAssembly package.
What people say
The pitch that made whisper.cpp a fixture of Hacker News threads still holds. You can clone the repo, download a single ggml model file, and get offline speech-to-text that sidesteps Python environments, CUDA installs, and API bills. Georgi Gerganov's C/C++ port runs the same OpenAI Whisper weights on hardware the original never reached — users show it transcribing on an iPhone 13, a Raspberry Pi 4 and inside the browser through WebAssembly. On Apple Silicon, the advantages are stark. Metal runs inference on the GPU, and a Core ML conversion can push the encoder onto the Neural Engine for a speed-up of roughly threefold over CPU-only runs. Quantization matters too — the q5_0 variant cuts large-v3 from 2.9 GiB on disk to 1.1 GiB, and large-v3-turbo-q5_0 squeezes into 547 MiB.
Users consistently run into the same set of problems. On an NVIDIA card, faster-whisper is quicker for the same accuracy, so GPU-heavy batch pipelines tend to go there. Whisper's own habits carry over: silence at the end of a recording can produce confidently fabricated text — a long-standing open issue — and long recordings with sparse speech invite repetition loops. Speaker diarization exists only as the experimental tinydiarize small.en-tdrz model, which marks speaker turns rather than naming speakers, so people who need real diarization bolt on WhisperX or pyannote. Aggressive quantization costs accuracy on noisy audio, and the engine is inference-only — no fine-tuning.
The project continues to evolve quickly. It moved from Gerganov's personal account to the ggml-org organization, shipped five releases between late May and mid-June 2026, and added native Silero voice-activity detection, which trims both hallucinations and processing time by skipping non-speech. A deep bindings list — Rust, Go, Ruby, Java, .NET, React Native, Unity — means whisper.cpp is the engine inside a lot of dictation and note-taking apps whose users have never heard of it.
Summary of public user & expert reviews, compiled by RECATOOLS.
About this listing
This entry was compiled from publicly available data including whisper.cpp's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with whisper.cpp unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to whisper.cpp directly →
Spotted something out of date? Suggest an update →
Alternatives to whisper.cpp
More in Video & Audio
