Transcribe Audio Locally with whisper.cpp, No Cloud
whisper.cpp runs OpenAI's Whisper speech recognition on your own hardware, so recordings of meetings and calls stay on the machine. This tutorial takes you from clone to a finished transcript using only the command line.

What you will have and what you need
At the end you will have a locally built whisper-cli program, a downloaded model and a text transcript of an audio file. None of the steps below sends audio to an external service. The tool is a plain C/C++ implementation without dependencies and supports CPU-only inference, so no graphics card is required.
You need git, CMake, a C/C++ compiler and ffmpeg for converting recordings. The model download needs internet access once, and afterwards the machine can work offline. The README lists disk and memory needs per model: base takes 142 MiB on disk and about 388 MB of memory, small takes 466 MiB and about 852 MB, and large takes 2.9 GiB and about 3.9 GB.
Clone the repository, download a model and build
First fetch the source code named in the whisper.cpp README and change into its folder. The two commands below do exactly that.
git clone https://github.com/ggml-org/whisper.cpp.git
cd whisper.cppNext comes the model. The script below downloads the multilingual base model in the ggml format that whisper.cpp uses. The README names further sizes, from tiny up to large-v3-turbo, and the models with an .en suffix handle English only.
sh ./models/download-ggml-model.sh baseThe two CMake commands below configure the project and compile it in Release mode. The resulting binary lands in build/bin.
cmake -B build
cmake --build build -j --config ReleaseConvert a recording and transcribe it
The README states that whisper-cli currently runs only with 16-bit WAV files. The ffmpeg call below turns an MP3 into a WAV with a 16000 Hz sample rate, one channel and 16-bit samples. Replace the file names with your own recording.
ffmpeg -i besprechung.mp3 -ar 16000 -ac 1 -c:a pcm_s16le besprechung.wavNow run the transcription. The command below passes the model file with -m and the audio file with -f, and it prints the recognised text with timestamps in the terminal.
./build/bin/whisper-cli -m models/ggml-base.bin -f besprechung.wavFor all options, the README points to the help output of the tool, which you get with ./build/bin/whisper-cli -h. The documentation also describes an optional voice activity detection with the --vad argument, which passes only detected speech segments to Whisper.
How to check that the transcription works
Before you try real recordings, test the setup with the sample file that ships with the repository. The commands below download the English-only base.en model and transcribe samples/jfk.wav, a short speech clip of about 11 seconds. If the spoken sentence appears as text with timestamps, build, model and tool work together.
sh ./models/download-ggml-model.sh base.en
./build/bin/whisper-cli -m models/ggml-base.en.bin -f samples/jfk.wavAfterwards run your own file and read the result against the recording. Names, product terms and numbers deserve a manual check, because speech recognition output is a draft and never a finished record.
Typical pitfalls according to the documentation
The most common stumbling block is the input format. Because whisper-cli accepts only 16-bit WAV files, an MP3 or M4A file needs the ffmpeg conversion first. The README also mentions an optional build with WHISPER_COMMON_FFMPEG that lets some examples decode other formats directly.
- A model downloaded as base is called ggml-base.bin, while base.en gives ggml-base.en.bin, so the file name in the command must match.
- English-only models carry the .en suffix, so German recordings need the multilingual variants such as base or small.
- Hardware acceleration needs its own CMake flag at build time, for example -DGGML_CUDA=1 for NVIDIA cards or -DGGML_VULKAN=1 for Vulkan.
- The README lists inference as the only limitation, so the tool transcribes and does not train models.
The documentation does not say which model size gives the best accuracy for German meeting audio. Compare a few sizes on your own recordings, since larger models need more disk space and memory.
Use in a company: operation and data protection
Meeting recordings often contain personal data, and a transcription that runs on your own server or workstation keeps the audio inside your network. Companies still tend to check who may record, how long audio and transcripts are stored and whether a works council (Betriebsrat) must be involved. The overview of GDPR and AI at work covers those questions in more detail. For holders of professional secrets under section 203 of the German Criminal Code, local processing is a frequent starting point, as the pages for law firms and tax advisers explain.
Someone in IT should own the installation, update the build when the project releases new versions and decide which model sizes are approved. The README also describes Docker images and a whisper-server with an HTTP endpoint, which suit a shared internal service. More background is on the page about local AI.
souveraen.ai is an AI platform for company knowledge from clouditiv GmbH in Auerbach. It runs either hosted in Germany or on-premise, and the on-premise option is Private Spark on an NVIDIA DGX Spark in the company network.
Sources
- whisper.cpp README, whisper.cpp


