edytlab
User Guide

Stem Separation and Transcription

Separating stems, transcribing speech, exporting captions and markers, and finding who speaks when, with models you download.

On this page4 topics

Stem Separation

Ask the agent separate the stems on track 1 and it splits the track into vocals, drums, bass and other with HT-Demucs, on your computer. No audio is sent anywhere, and separating makes no network request. The result is four WAV files in the project folder; the session itself does not change. To work with a stem, ask the agent to load it, and it becomes a track like any other. Stems are estimates, not the original recordings: expect some of one instrument to leak into another stem.

The model is not installed with the app. Open Settings → Account and press Download beside the Stem model row (about 171 MB in two files; see LLM Provider and Model). Nothing downloads unless you press it, and the agent cannot download it. The files come from huggingface.co, which sees your IP address and that edytlab is asking, and nothing about your audio.

  • What it works on. A track as it sits on the timeline, or an audio file by path. Stems keep the sample rate and length of the audio, and a mono file gives mono stems.
  • Time and memory. It runs on the CPU and takes about half the length of the audio. It needs about 4 GB of free memory; a computer with less cannot run it, and the result says so. The app may hold on to some of that memory until you close it.
  • Approval, progress and Cancel. A run writes about 40 MB for every minute of stereo audio, so Plan first holds it for your approval. A progress strip shows the work, and Cancel stops it and leaves nothing behind.
  • Disk. The stems are kept in the project’s .audiograph folder, up to about 1 GB: the ones used longest ago go first, and the set just made is kept even when it alone is larger. Asking again for audio whose stems are still there is instant. A stem you have loaded is copied into the project, so clearing the cache never breaks a track.
  • What reaches your AI provider. The audio never leaves your computer. The paths of the four stems come back as a tool result, so they go to the provider you chose, as text, and stay on your computer with Ollama.

HT-Demucs is Meta’s (MIT licence) and the files are Intel’s ONNX export of it (also marked MIT); the credits are on the Terms page.

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.

Transcript Export and Markers

After transcribing, ask export the transcript to captions.srt to write an SRT file, or name vtt for WebVTT or txt for plain text. The export_transcript tool needs out_path, relative to the project folder or absolute. It writes the file atomically, protects project audio from being overwritten, and changes no audio or session history.

Captions split at sentence-ending punctuation and at a pause of 0.75 seconds by default, and normally stay within six seconds and two lines of 42 Unicode scalar values each (not display columns). Ask for different pause_sec (0 or more), max_cue_sec (greater than 0) or max_chars_per_line (an integer from 1 to 200). A single stored word may need more lines or exceed the duration limit, because the export keeps its recorded times. TXT keeps each segment on one unwrapped line. All three formats use UTF-8 and LF line endings. WebVTT escapes markup characters; SRT keeps raw text, which readers may interpret as markup.

Ask add a marker at each transcript sentence to use transcript_to_markers. It adds point markers in one undoable step, keeps existing annotations, and changes no audio. Choose mode: segment (the default caption grouping) or mode: sentence (sentence-ending punctuation). Labels use the first six words, capped at 80 Unicode characters. Running it again skips markers that already exist (same name, within 1 ms).

Both tools use the stored word times. A track transcript keeps session seconds, and its words follow edits that move, cut or stretch the track; align_to_beat leaves them stale, so transcribe again after it (both tools say so in their results). A file transcript keeps file seconds; an older transcript may have an unknown source. Results report that time basis and source, with a stable track ID and its current index when available. For markers from a file or an unknown source, explicitly supply a signed offset_sec to align those times to the timeline, even if the offset is 0. Every resulting time must be finite and at least 0. A removed source track is refused even with an offset; transcribe an existing track again.

Who Speaks When

Ask the agent who speaks when on track 1 and it runs the speaker model on your computer (diarize) and answers with the speaker turns: start and end times, numbered speaker_1, speaker_2 in the order they first speak. It changes nothing in the session. To give each voice its own track, ask it to split by speaker; that step (split_by_speaker) is an edit like any other and waits for your approval.

The model is not installed with the app. Open Settings → Account and press Download beside the Speaker model row (about 33 MB in two files; see LLM Provider and Model). Nothing downloads unless you press it, and the agent cannot download it. The files come from huggingface.co, which sees your IP address and that edytlab is asking, and nothing about your audio.

  • What it works on. A track as it sits on the timeline (turn times are session seconds), or an audio file by path (times are the file’s). If you know how many people speak, say so; otherwise the count is estimated.
  • Limits. Voices that sound alike can merge into one speaker, and where people talk over each other the audio goes to one of them. Check the turns before you split. A very long conversation is listed only in part, and the result says so.
  • What reaches your AI provider. The audio never leaves your computer. The turns (times and speaker numbers) come back as a tool result, so they go to the provider you chose, and stay on your computer with Ollama.
  • Speed and Cancel. A progress strip shows the work, and Cancel stops it without changing the session.

The segmentation network is pyannote’s (MIT licence, © CNRS) and the voiceprint network is WeSpeaker’s (CC BY 4.0); the credits are on the Terms page.