Batch Baking¶
Baking a folder of clips in one go, with their transcripts found for them.
Batch is the same bake as single clip, run over a folder. Everything about strategy, language, recognizer and output works identically. What changes is how the baker finds clips and pairs each one with its words.
Switch to it with Mode in the Audio section.
Folder In, Cues Out¶
Audio Path is the folder holding the recordings. WAV, OGG, MP3 and FLAC are all read. If a clip exists in more than one format, one is chosen and the others ignored, so a folder that has been converted at some point does not bake twice.
Then the baker has to work out which words belong to which clip, and there are two ways to tell it.
Transcript Sources¶
Each recording sits next to a text file with the same name. barks_01.wav pairs with barks_01.txt.
This is the shape most speech datasets already use, so a folder exported from almost anywhere works without rearranging. It is also the easiest to produce by hand, and the easiest to edit later, since one line of dialogue is one file you can open.
One spreadsheet holds every line, and the audio folder holds every recording. Rows are matched to files by clip name.
The columns are clip_id, language and transcript. A header row is required, the column order does not matter, and extra columns are ignored, so a sheet your team already uses for something else can usually be pointed at directly.
Because each row carries its own language, this is the source to use for a mixed-language batch.
clip_id,language,transcript,speaker,scene
barks_01,en,Watch your step down there.,Mara,cellar
barks_02,en,<Worried>I would not go in alone.</Worried>,Mara,cellar
greet_ko_01,ko,반갑습니다,Jun,market
barks_01.wav and the rest are found in the audio folder by those ids. speaker and scene are along for the ride and the baker ignores them. Transcript tags work here exactly as they do anywhere else, so the second row bakes an emotion cue alongside its mouth movement.
Articy XML is a third source, available when the Articy integration is enabled.
What Matched, and What Did Not¶
Neither pairing is guaranteed to be complete, so the baker tells you where it stands before you commit to a run. A coverage line reports how many discovered clips have transcripts and how many do not.
The two failures are treated differently on purpose:
- A recording with no transcript is skipped quietly. That is the normal state of a folder you have not written transcripts for yet, and it is what the coverage count is for. It is a stage in a workflow, not a mistake.
- A transcript with no recording warns. That one usually is a mistake, a renamed file or a typo, and staying quiet about it would let a line silently never bake.
Nothing in either case stops the run. Everything that pairs gets baked.
Generating What Is Missing¶
The coverage count is not just a report. If clips are missing transcripts, the baker will write them from the audio, in one pass over the batch, in whichever source format you selected. So an unpaired folder becomes a paired one without leaving the window.
It works in the other direction too. A batch with transcripts and no audio can have the audio synthesized.
Language detection is per clip here, so a folder of mixed-language recordings transcribes without being sorted first. See Generating a Transcript.
Running the Batch¶
Three run modes, chosen in Setup ▸ Performance. The current one is named next to a coloured dot above the bake button, and that label is clickable if you want to change it.
| Mode | What it costs you |
|---|---|
| Sequential | The editor stays responsive and the run can be cancelled. It pauses between clips while the editor is in the background and resumes when you come back to it. |
| Synchronous | The editor is frozen for the whole batch and Cancel will not respond. In exchange the run keeps going while the editor is in the background. |
| Threaded | Clips bake on a background thread and the editor stays interactive. |
The honest trade is between being able to use the editor and being able to walk away from it. Sequential lets you keep working but stalls when you alt-tab. Synchronous is the one to pick for a long run you intend to leave alone.
Threaded Is Experimental and Usually Does Not Work
On most projects only the first clip succeeds and the rest fail. It is exposed because it is the path this will eventually take, and it is marked so nobody discovers that the hard way. Use Sequential or Synchronous unless you are deliberately testing it.
After the Run¶
The result area reports what was baked and what was skipped.
Show In and Add To work as they do for a single clip, with one difference: they act on the whole output folder rather than one file. Add To hands the folder to the registry, which groups the outputs by clip name for you, so registering a hundred clips is the same amount of work as registering one.
Check the Output Path Before a Large Run
An empty Output Path means each cue file is written beside its own audio. For a batch drawn from several folders that scatters the results back across all of them, which may well be what you want. If it is not, set a folder and everything gathers in one place. See Output.