A podcast from a recorded call is the classic case. Zoom, Meet and Teams save a video file even when the conversation is the whole point, and a podcast host or an audio editor wants sound only. Extract the track, open the MP3 in your editor, trim the small talk and the episode is most of the way there.
A voice-over is another. You narrated a tutorial or a product video and now need the same narration for a shorter cut, a slideshow or a different language version to time against. Taking the audio from the finished export is quicker than hunting for the project file.
The third is simply size. A twenty-minute video of someone talking can be too big for e-mail and awkward on a messaging app, while its audio is a file of 20 to 30 MB that sends anywhere and plays on anything.
The result is the complete soundtrack as one file. Everything in the mix comes along together: voices, music and background noise are a single track in the video, and extraction cannot pull them apart. Separating a voice from the music under it is a different job that needs a stem-separation tool.
With Original audio the track is copied, not converted. Phones, cameras and screen recorders store sound as AAC almost without exception, and that track becomes an M4A file bit for bit, so the quality is exactly what was recorded and the work is only rewrapping, which is why it is so fast. Opus, Vorbis, MP3 and FLAC tracks are copied the same way. A track in a format audio players rarely open, such as uncompressed PCM from some cameras, is saved as MP3 instead. Choose MP3 yourself when a program insists on it; converting one lossy format to another costs a little detail, inaudible for speech at sensible bitrates. If you need only a part of the soundtrack, extract it first and then use the audio cutter to keep the range you want.
The recordings people extract audio from are often the ones they should not hand to an unknown website: client calls, job interviews, internal meetings, unreleased footage. Here the page loads ffmpeg as a WebAssembly program and runs it in your browser. The video is read from your disk and the MP3 is written back to it; no copy passes through a server.
The first use downloads the engine, about 30 MB, and the browser caches it afterwards. How long a conversion takes depends on your device, and on a laptop it is normally far quicker than playing the video. Very large files, in the range of several gigabytes, may be too much for a phone and belong on a computer. Videos with DRM, such as purchased films, cannot be processed.
Show notes for the podcast, minutes for the call, quotes for an article: much extracted audio ends up being typed out. The "Transcribe this file" button that appears after the extraction sends the audio to Transcrify's free transcription. For a transcript, the audio is all that is needed: the speech model hears exactly the same thing as in the video, so the text is identical, and the upload is a fraction of the size.
Be aware that this is the one step where the file leaves your device. The audio is uploaded to servers Transcrify runs itself and transcribed there, not by an outside speech service. 10 recordings are free, each up to 3 hours and 5 hours in total across guest use and a free account. A guest gets a timestamped preview, the full text for short recordings and about the first ten minutes of long ones, and a free account keeps the entire transcript.