How to Transcribe Audio Offline for Free With whisper.cpp
Run the free whisper.cpp tool locally to turn audio into editable text while understanding hardware, models and accuracy limits.
Before you begin
Interviews, lectures, meeting notes and recorded voice memos are easier to search when converted into text. Cloud transcription services may require uploading audio or paying for usage. whisper.cpp is a free open-source implementation of Whisper inference that can run locally on supported computers after downloading the software and a model. It is more technical than a one-click web app, but it offers useful control over where processing happens. Transcripts still need human correction.
1Check the official project and your hardware
Open github.com/ggml-org/whisper.cpp and read the Quick start and platform requirements. The project supports several operating systems and CPU or accelerator paths, but setup details differ. Model sizes vary substantially in download size and memory needs. Begin with a smaller model for a short clip rather than attempting the largest one immediately. If you are using a managed computer, obtain permission to install developer tools and download model weights.
2Install the application from source or a trusted release
Follow the repository's documented installation or build instructions rather than downloading an unknown packaged executable from a file-sharing site. The official example uses a Git clone, CMake configuration and a build step to create the whisper-cli program. On Windows, command paths may differ from Linux and macOS examples. Keep notes on the version and build options you used so later troubleshooting is reproducible.
3Download a suitable speech recognition model
The project's documentation includes scripts for obtaining GGML-format models. For English-only recordings, a model named with an .en suffix can be a reasonable starting option. For other languages, use a multilingual model compatible with the software and consult the language options. Models are separate files and their size is not an indication of guaranteed accuracy. Small models are often faster but may struggle with difficult accents, background noise or technical vocabulary.
4Prepare audio in a supported format
The official command-line example expects compatible WAV audio, with a documented conversion example using FFmpeg to create mono 16-bit PCM audio at 16 kHz. For example, a video or MP3 recording may need a separate format-conversion step first. Save converted audio in a clearly labeled folder and retain the original recording. Obtain consent when recording or processing conversations, and follow applicable workplace or privacy obligations.
5Run a short transcription and save the text
Use the whisper-cli executable with the correct model path and the prepared audio file, following the exact syntax for your installed version. The repository shows a command with an input file such as samples/jfk.wav, and its command-line help lists optional switches. For a long recording, test sixty seconds first to assess speed and word accuracy. Save the resulting transcript to a text or subtitle-compatible format only using options confirmed by the installed tool.
6Proofread names, numbers and quotations
Automatic speech recognition can omit words, join speakers incorrectly, hallucinate phrases in silence or change a number. Listen to the audio again while checking names, dates, prices and technical terms. If you need subtitles, adjust sentence boundaries and timing so captions remain readable. Keep the original file and a corrected transcript, with labels indicating what was machine-generated and human-checked. For legal or medical records, do not treat unreviewed transcription as authoritative.
Frequently asked questions
Can whisper.cpp work without internet?
After installation and model download, it can perform local inference without uploading audio to a cloud speech API.
Does free transcription mean unlimited speed?
No. Processing time depends on your computer, model choice and recording length.
Official sources and further reading
Interface labels, availability and policies may change. Confirm current details with the relevant service before taking action.