Almost nobody runs a single audio operation in isolation. The real work is a sequence: pull the audio out of the video, separate the vocal, clean up the noise, transcribe it. Then you do the exact same sequence again on the next file. And the next. And the next.
That repetition is where two things go wrong: you re-assemble the same steps by hand every time, and you have no clean way to run them across a whole folder at once. Recipes and bulk execution are our answer to both.
Recipes: design the pipeline once
A recipe is a saved, multi-step pipeline. You compose the steps once, where each step is a real capability like stem separation, noise reduction, or transcription. Save it, and from then on you invoke the whole thing by ID. The pipeline becomes a reusable object instead of a sequence you rebuild from memory.
You can create and save your own recipes, keep works-in-progress as drafts, and browse a growing public library of recipes we curate. (That public library is still being populated, so it grows over time, but creating and running your own recipes is fully live today.)
Bulk: run it 500 at a time
Once a pipeline is a recipe, running it on one file or five hundred is the same shape of request. The batch API is the developer surface built for exactly this, and it's the one we've put through GA validation, because "bulk" is only useful if it's trustworthy at scale. What that gets you:
- Up to 500 items per batch. Ask for more and you get a clear, coded error, not a silent truncation.
- Fair concurrency. One giant batch can't starve everyone else's work. Over-cap items are deferred and eventually dispatched. None are quietly dropped.
- Retry by lineage. Failed items can be requeued without re-running the ones that already succeeded.
- Cancel mid-flight. Stop a batch that's still running; work already in progress finishes cleanly.
- Honest, coded errors. Every failure mode has a stable code you can branch on, instead of a mystery 500.
Why this matters
The difference between a toy and infrastructure is what happens on the hundredth file. A toy makes you click through it a hundred times. Infrastructure lets you describe the work once and trust the system to run it a hundred or five hundred times, fairly, with a real story for what happens when something goes wrong.
Design the pipeline once. Run it at the scale your actual work demands.
Recipes and the batch API (500-item cap, fair concurrency, retry, cancel, coded errors) are live; the batch surface is GA-validated. See Developers.