Whisper

Whisper is OpenAI's open-source speech recognition model, released with weights and code under the permissive MIT licence. Instead of paying per minute for a transcription API, you run Whisper locally on your own machine and turn audio into text — in roughly a hundred languages, with translation into English built in.

It is the engine behind a large share of free transcription tools. Run directly, it takes files from the command line and produces transcripts with punctuation and paragraphing. Wrappers add drag-and-drop interfaces, and it pairs naturally with local model runners for asking questions about a transcript afterwards. Because the model runs offline, recordings never leave your computer.

It fits podcasters and students transcribing lectures, journalists working with interviews under source-protection constraints, developers adding speech-to-text to their own projects without per-minute fees, and anyone with sensitive audio that should not be uploaded anywhere.

To begin, install it from the GitHub project (Python and ffmpeg required) and run whisper yourfile.mp3 from the terminal — or pick a graphical Whisper app from Hugging Face if you prefer not to touch a terminal. Start with the base or small model; larger ones are more accurate but slower.

The honest limits. It is a model, not a product: the reference implementation is a command-line tool, so comfortable setups need either the terminal or a third-party wrapper. Large models need meaningful RAM or GPU. Accuracy drops on heavy accents, crosstalk, and poor recordings, and there is no built-in speaker identification. Model downloads are large.