Whether Whisper is worth installing depends on how much you need to transcribe and what hardware you have. It is an open-source speech recognition model OpenAI released in 2022: audio is transcribed on your own computer, with no per-minute charge. People with only the occasional recording are better off with a ready-made transcription service; people who process meetings, interviews, or podcasts every week save real money and waiting time by installing it once.
Official repository information
The official repository is on GitHub, the organization name is openai, the project name is whisper, and the code and model weights are released under the MIT license. The models come in six sizes — tiny, base, small, medium, large, and turbo — and the reference VRAM figures are roughly 1 GB each for tiny and base, about 2 GB for small, about 5 GB for medium, about 6 GB for turbo, and about 10 GB for large, with speed measured against large and tiny the fastest. Real usage varies with the runtime, so check your graphics card against this table first; with no GPU at all, you can still start with tiny or base.
Deployment itself is not complicated — the time goes into choosing a runtime. The original version runs with Python plus FFmpeg, faster-whisper is the choice for speed, and whisper.cpp is the choice for running on a CPU or Apple Silicon. Models download on first run and can then work fully offline. Having no interface is part of its nature: it only turns audio into text, so dictation, live captions, and automatic segmentation all need other tools connected to it — do not expect to open a meeting with it straight after installing.
Three real pitfalls, know them before you install
- Silent passages invite hallucination. Long silences or background noise can lead the model to write words nobody said; adding voice activity detection in front and cutting long audio into segments before transcribing reduces this noticeably.
- Proper nouns and people's names will still be wrong. Names, places, and product names need a human check, and in important settings never send the transcript out without checking — numbers and times especially need to be verified against the original audio.
- Timestamps drift on long audio. On recordings of two or three hours, segment boundaries and speaker changes are where alignment slips most, so it works as a draft transcript, but not as a final set of minutes.
It suits people who transcribe on a fixed weekly schedule and care that audio does not leave their machine; it does not suit people who want to click and go, or whose computer is too old to run even small. For the latter, first try small on a short recording — if it runs, then decide whether to move up to large.
