whisper.cpp — Local Speech-to-Text in Pure C/C++

What it is

whisper.cpp is a high-performance C/C++ port of OpenAI's Whisper speech recognition model by Georgi Gerganov (creator of llama.cpp). It runs entirely locally with zero dependencies: no Python, no PyTorch, no internet connection needed.

The key advantage: it runs efficiently on CPU. Apple Silicon gets 4-8x speedup via Core ML and Metal. NVIDIA GPUs work via CUDA. Even a Raspberry Pi can transcribe audio. Real-time streaming transcription works on modern laptops.

How it saves time or tokens

whisper.cpp provides speech-to-text without cloud API costs or latency. Traditional Whisper requires Python, PyTorch, and ideally a GPU. whisper.cpp runs on any hardware with a single binary. For privacy-sensitive applications, all processing stays on-device. The tiny model (75 MB) transcribes at 32x real-time on CPU, making it practical for batch processing of audio archives.

How to use

1. Clone, build, and download a model:

git clone https://github.com/ggerganov/whisper.cpp
cd whisper.cpp
cmake -B build
cmake --build build --config Release
bash models/download-ggml-model.sh base.en

1. Transcribe an audio file:

./build/bin/whisper-cli -m models/ggml-base.en.bin -f samples/jfk.wav

1. Real-time microphone transcription:

./build/bin/whisper-stream -m models/ggml-base.en.bin
# Speak into your microphone -- text appears in real time

Example

Model size comparison for different use cases:

| Model  | Disk    | RAM     | Speed (CPU)    | Quality          |
|--------|---------|---------|----------------|------------------|
| tiny   | 75 MB   | ~390 MB | ~32x real-time | Good for drafts  |
| base   | 142 MB  | ~500 MB | ~16x real-time | Solid accuracy   |
| small  | 466 MB  | ~1 GB   | ~6x real-time  | Good quality     |
| medium | 1.5 GB  | ~2.6 GB | ~2x real-time  | High quality     |
| large  | 2.9 GB  | ~4.7 GB | ~1x real-time  | Best quality     |

# Output formats
./build/bin/whisper-cli -m models/ggml-base.en.bin -f audio.wav -otxt   # Plain text
./build/bin/whisper-cli -m models/ggml-base.en.bin -f audio.wav -osrt   # SRT subtitles
./build/bin/whisper-cli -m models/ggml-base.en.bin -f audio.wav -ovtt   # VTT subtitles
./build/bin/whisper-cli -m models/ggml-base.en.bin -f audio.wav -ojson  # JSON with timestamps

Related on TokRepo

• AI tools for voice — More speech and voice tools on TokRepo.
• Local LLM tools — Browse local AI inference tools.

Common pitfalls

• Using the large model on hardware without a GPU leads to very slow transcription. Start with base or small for CPU-only setups.
• Audio files must be 16kHz 16-bit mono WAV. Convert other formats with ffmpeg before processing.
• Real-time streaming requires a low-latency audio capture setup. Ensure your microphone input is configured correctly for the whisper-stream binary.