Run Whisper Locally on Mac: A File Transcription Guide

A separate local transcription workflow

Running Whisper locally means that a model on your Mac processes the audio. It is a separate workflow from Sonicribe Cloud, which sends audio to online services and has local speech processing disabled.

This guide uses whisper.cpp for a basic file-transcription setup. Begin with a public or invented recording so you can check the workflow before using personal material. You will need internet for the initial software and model downloads; the subsequent local transcription step can run without it.

Prepare your Mac

Have Git, CMake, and the developer tools needed to compile the project available. Allow space for the source, build files, model, audio, and output. If you are unfamiliar with these tools, use the project's setup documentation before proceeding.

Choose a folder for the project and keep your recordings in a separate folder you can find again. Check whether either location syncs to another service if you want the resulting files to stay on your Mac.

Build whisper.cpp and try a sample

The official whisper.cpp quick start documents cloning, downloading a model, compiling, and running the command-line example. Follow the current instructions if the project changes.

Run these commands in Terminal:

git clone https://github.com/ggml-org/whisper.cpp.git
cd whisper.cpp
sh ./models/download-ggml-model.sh base.en
cmake -B build
cmake --build build -j --config Release
./build/bin/whisper-cli -m models/ggml-base.en.bin -f samples/jfk.wav

This first example uses an English-only model and the repository's sample audio. Check that the model file exists and the command produces text before trying your own recording. The setup downloads resources; the transcription command uses the downloaded model locally.

Prepare your own recording

The documented command-line example expects 16-bit WAV input. If your audio is in another format, the project documents conversion with FFmpeg:

ffmpeg -i input.mp3 -ar 16000 -ac 1 -c:a pcm_s16le output.wav
./build/bin/whisper-cli -m models/ggml-base.en.bin -f output.wav

Replace the input filename with your recording, and choose an output filename that will not overwrite a file you need. FFmpeg must be installed before using the conversion command. Keep the original audio until you have reviewed the transcript.

Inspect the available command options locally:

./build/bin/whisper-cli -h

Use the help output to choose language and export options for your installed version rather than copying flags from a different version.

Verify offline operation and review the text

Once the model and software are installed, disconnect from the network and transcribe the same sample again. Repeat with your own short recording. If you add a graphical wrapper, plugin, or summarizer, test that component's processing boundary separately.

Compare the transcript with the recording. Review names, numbers, dates, negations, and technical terminology. Save the corrected result in a format you can open in your destination app.

A file-transcription command is not automatic dictation at the cursor. If you need live microphone capture or text insertion, plan those as additional workflow steps rather than expecting this example to supply them.

Choose a model through a small comparison

Begin with a model that runs comfortably on your device, then evaluate another model on the same sample if you need a different result. Compare turnaround and editing effort, not only the apparent fluency of the text.

For non-English audio, choose a multilingual model supported by the local project and follow its model setup documentation. In Sonicribe, choose the spoken language in Settings → Listening → Speech language and the online speech model in your style’s Voice section.

If a model fails to load, check the path and completed download. If transcription fails on your audio, check its format. If processing is impractical, try a smaller model and record the trade-off in corrections.

Sonicribe Cloud is an online alternative

Sonicribe requires an Apple Silicon Mac, a Google account, and internet. Windows and Intel Macs are unsupported. It does not download or run the local model used in this guide.

Eligible new accounts get 100 free minutes with Qwen3-ASR, with no time limit. Pro is US$7/month for up to about 2,500 minutes a month; Max is US$20/month for up to about 10,700 minutes a month with Qwen3-ASR. Optional online formatting shares the speech balance. Unused paid credits reset monthly, and recording pauses at zero credits.

Review pricing and the privacy policy if online processing fits your needs. For local processing, continue with the separate setup above and review where your recordings, transcripts, and exports are stored.