Online MFCC extractor memory growth in Kaldi 2024‑06 stable branch

Symptom

Continuous streaming with the online MFCC extractor shows a gradual increase in resident memory over multi‑hour runs, even after the buffer‑reset patch (commit 3b7a9c2) was merged into the 2024‑06 stable branch.

Goal

Confirm whether the leak is fully resolved in the current stable release and determine the appropriate default for the --apply-preemphasis flag, which influences both robustness and computational load.

Constraints

• Variable‑length utterances processed in a single process.
• Memory must remain bounded without manual intervention.
• Downstream acoustic model performance should not degrade.

Open questions

1. Does the latest stable branch still exhibit unbounded memory growth during extended streaming?
2. What default setting for --apply-preemphasis balances accuracy and resource usage?
3. Are additional buffer‑management steps required beyond the 2024‑06 patch?