Local Whisper Transcription: Running Speech-to-Text on Your Own Computer
Whisper models can transcribe audio on an ordinary computer, offline once downloaded. Which model size to pick, what it needs, and where it falls short.
The Scriptorium team ·
You can run Whisper, OpenAI's open speech-recognition model family, on your own computer, with no per-minute fee and no recording leaving the machine. The trade-off is speed: on a processor without a GPU, bigger models run slowly.
What you need
A computer with enough memory for the model you choose, and a one-time download of the model weights. After the download, transcription needs no internet connection. Scriptorium runs the models through faster-whisper, an efficient implementation that works on an ordinary CPU.
Choosing a model size
- tiny (about 75 MB): fastest, rough-draft quality.
- base (about 145 MB): fast, fine for clear speech.
- small (about 480 MB): a good balance, and the usual starting point.
- medium (about 1.5 GB): more accurate, noticeably slower.
- large-v3-turbo (about 1.6 GB): near-best accuracy, faster than large.
- large-v3 (about 3 GB): best accuracy, slow on a CPU.
Start with small. Move up only if names, accents or technical terms come out wrong often enough to justify the wait.
What local transcription does not fix
Running locally changes where the work happens, not how well the model hears. A noisy room, overlapping speakers and strong accents are just as hard on your own computer as in the cloud. Whisper also needs help with speaker labels, which is a separate step.
Using it without the command line
If you would rather not assemble the pieces yourself, self-hosted transcription software wraps the model in a web interface with uploads, speaker labels, editing, search and export.