Stem Separation and Transcription

Stem Separation

Not available in this build. Stem separation is meant to split a track into vocals, drums, bass and other with Demucs, on your device, but inference ships as a stub, so asking the agent to separate the stems on track 1 returns an error. There is no stem model to download and no environment variable that changes this: the decoder is what is missing, not the model file.

Transcription

Ask the agent to transcribe track 1 and it writes what was said, with the start and end time of every word, as a new version of the session. It runs on your CPU with the Whisper base speech model. No audio is sent anywhere, and transcribing makes no network request.

The model is not installed with the app. The first time, open Settings → Account and press Download beside the Speech model row (about 139 MB; see LLM Provider and Model). Until it says Ready, a transcript request shows a card saying what is missing, with a button for Settings. The agent cannot download the model: only your click does. A build from source also needs the ONNX Runtime library, which the installers carry.

• What it transcribes. A track as it sits on the timeline (word times are session seconds), or an audio file outside the session, by path (times are the file’s). A new transcription replaces the earlier one.
• Language. The model detects it, or ask for one by two-letter code (en, de, fr, es). The model is multilingual Whisper base, checked only on English: a short synthetic recording and about 16 minutes of audiobook speech, where 4% to 7% of the words were wrong. Read the words before you cut on them.
• Speed and Cancel. A progress strip counts the 30-second windows, and Cancel stops it without changing the session. Speed depends on your CPU: on our test machine seven seconds of audio took about three seconds the first time, and the same audio again is answered from a cache. The whole recording is decoded into memory first: about 1.3 GB for an hour of 44.1 kHz stereo.
• The Transcript tab. It shows the words with their times. Select words and press Delete to cut that audio and close the gap. The tab cuts only the track the transcript is of, and says why when it will not.
• Editing by words. Ask the agent to cut a passage (cut_words), find and remove fillers such as um and uh (remove_fillers, which changes nothing until you say to apply it), or select a phrase (select_region). Each edit is one undoable step.
• What reaches your AI provider. The words come back as a tool result, so the text of what was said (a bounded amount per call, never the audio) goes to the provider you chose, and stays on your computer with Ollama.

The rest of the editor works without a transcript or the model.