Just saw the announcement for the new Stem Splitter feature in Suno. As someone who often tinkers with audio processing pipelines (FFmpeg filters, anyone? 😅), I'm naturally curious about the technical implementation and practical results.
Has anyone put it through its paces yet? I'm particularly interested in:
* **Separation quality:** How clean is the isolation? Does it introduce artifacts, especially on complex tracks with layered instruments?
* **Latency/Overhead:** For batch processing, is it fast? Does it feel like a quick `sox` chain or something heavier?
* **Formats & API:** Does it work on various input formats (e.g., `.flac`, `.mp3`)? Is there a local/CLI option, or is it purely cloud/API-driven? The latter would add network RTT to the "processing time."
I did a quick test with a dense rock track. Using the web interface, the split into vocals, drums, bass, and other was... decent.
```bash
# Not actual Suno CLI, but this is the kind of workflow I'm hoping for
$ suno-cli stem-split --input track.wav --stems vocals drums bass other
Processing 'track.wav'...
Output: 'track_vocals.wav', 'track_drums.wav', ...
```
The vocal isolation was good, but some synth elements bled into the "other" stem. The drum track was surprisingly clean, though.
How does it compare to other tools like `spleeter` or `demucs` in your experience? Is the quality worth the potential cost/credits for your use case? I'm thinking about automated pipelines where reliability and consistency matter more than one-off manual use.
System calls per second matter.