Whisper, here
The words are made on this computer by Whisper base, OpenAI's speech model, run through Transformers.js and ONNX Runtime in a worker. The first Transcribe says what it costs and downloads the model once, from this site: about 110 MB for English, and about 80 MB more for Any language. It is kept in the browser's private storage, so from then on transcribing works offline too, and Projects lists each model kept with what removing it frees. The first words take a moment, since the model loads and then reads the first half minute; after that they fill the panel as they come.
Words as a selection
Click a word to go there. Drag over words, or Shift+click, to select their time on that track; the selection moves each edge into the nearest quiet within 150 ms, since the words' times run a little late. The arrows walk the words and Enter goes to one, the word being played is marked, and the search finds a phrase among all the words.
Chapters
Press M at the playhead to put a marker there, and name it. In Export, "Write chapters from the markers" writes them into the file where podcast apps look for them: MP3, M4A, Ogg and FLAC; a WAV carries none.
Subtitles
Export writes the words as text, SRT or WebVTT, for every transcribed track or for the tracks chosen; when more than one track has words, the subtitles name the track. The words are part of the project and of its .oxyd file, about 550 KB an hour; they follow every edit of the audio under them and come back with Undo.
How wrong the words are
About one word in ten is wrong on ordinary speech (Whisper base averages 10.3% word errors on the Open ASR leaderboard), more with names, people talking over each other, music under the speech and strong accents. It leaves out "um" and "uh" by design, so it cannot find fillers. Its word times run late, mostly 0.05 to 0.2 s. One track is one speaker, and the words cannot be corrected yet. Speed: about 1.7 times real time once the model is loaded, measured on a 2017 fanless laptop; nothing has been measured on a recent laptop yet. English is the English-only model; Any language is the multilingual one, and has not been measured.
Questions
- Is the transcript made on my computer?
- Yes. The model runs in the browser, in a worker, and the words stay in the project.
- Which languages?
- English, with the English-only model, or Any language with the multilingual one, which has not been measured yet.
- Can I fix a wrong word?
- Not yet. The words cannot be corrected in the panel; a wrong word is still a place to click to.