We had a 96-minute Zoom recording of a client strategy session. It had business details, client names, and revenue numbers in it. Uploading it to a cloud transcription service was not an option. Most services store your audio, some use it to train their models, and the ones that promise privacy still charge by the minute at scale. We needed a private, free, and repeatable solution. That is what led us to whisper.cpp.

What whisper.cpp is and why it works offline

whisper.cpp is a C++ port of OpenAI's Whisper speech recognition model. The original Whisper model is open source. whisper.cpp takes that same model and runs it entirely on your local machine with no API calls, no internet connection required, and no audio leaving your computer. You download the model file once. After that, every transcription happens on your hardware. For a team handling client recordings with sensitive business information, this is the only architecture that makes sense.

The setup on Windows takes about 20 to 30 minutes the first time. You download the whisper.cpp source from GitHub, compile it using a build tool like CMake and a C++ compiler, then pull down the model weights. We used the medium model, which gives strong accuracy on English without being too large. Once compiled, the tool runs from the command line and outputs a clean SRT file alongside a plain text transcript. The first run confirmed it worked. Ninety-six minutes of audio processed in about 12 minutes on a standard Windows desktop PC.

96-minute client recording. 12 minutes to transcribe. Zero audio uploaded anywhere. Total cost: nothing.

The PowerShell wrapper that made it reusable

Running a compiled C++ tool from a specific folder every time is clunky. We built a small PowerShell wrapper script that accepts a file path as input, calls the whisper.cpp binary with the right flags, names the output file after the original recording, and drops the SRT and TXT files next to the source audio. The script lives in a shared folder on the BBC Drive. Any team member can call it from any machine with a single command: Transcribe-Audio "path\to\recording.mp4". No folder navigation. No remembering which flags to use. One command, files ready in minutes.

The SRT output is immediately usable in three ways. It can be added as captions to the video. It can be fed into a Balay ni Bruno & Co. AI as structured text so the AI can produce clean meeting notes, action items, and a summary. Or it gets archived alongside the recording as a permanent searchable record of what was said. For a 96-minute meeting, the AI-generated summary from the transcript takes about two minutes to produce. What used to be a manual hour of note-taking is now a 15-minute total process from raw recording to finished summary, done entirely inside our own systems.

Key Takeaways

  • whisper.cpp runs Whisper transcription locally with no API costs and no data leaving your machine.
  • A 96-minute recording transcribes in roughly 12 minutes on standard hardware.
  • The output SRT file plugs directly into caption tools, AI note takers, and video editors.
  • Local transcription is non-negotiable for client calls with confidential business information.