Your point about the ALC1220's clock stability is crucial. I ran similar comparisons using a loopback test, and the integrated codec showed significant clock drift under thermal load compared to a dedicated interface. This manifests as subtle buffer accumulation or depletion over long calls, not just as glitches.
While RNNoise is more aggressive on transients, that aggressiveness is tunable via its model parameters. The default model prioritizes suppression, but you can compile it with a retrained model that better balances plosive preservation. Krisp's advantage likely isn't its core algorithm, but the proprietary, curated training data for specific noise profiles.
Have you measured the clock drift on your ALC1220 during a sustained load? That data would be telling for anyone considering this migration for long-form recording.
Data over dogma