← All Projects

Voice Research

ongoing ML Research

Private voice and speech research. Details withheld by design.

PythonPyTorch

Activity Timeline

  • A controlled GPU fine-tuning run completed 2 of 5 folds within its thermal window.

    The two folds finished successfully before a Polaris ruling closed the window, leaving three of five outstanding.

    experiment
  • 22 files (153.6K words) incorrectly pass corpus guard; staged fix not yet applied.

    The narrative-analysis gate was letting 22 files and 153.6K words through as subject-authored. Patch 03 adds environment indirection for author attribution; combined, the patches resolve narrative levels from 17 to 32. Dispatch envelope is read-only, so fixes remain staged.

    bugfix
  • Storage crisis: 53.3 GB zero-replica files on orbis WSL identified; GCS backup initiated.

    71-file critical set (31.26 GiB) confirmed as zero off-host copies. GCS credentials verified live, parallel sync started. Owner ack required before irreversible WSL deletion proceeds.

    blockedmilestone
  • Diarization pipeline at 245× realtime; disk reserve decision queued.

    7 background runs in 74.5s GPU time, zero OOM. RTTM-to-candidate-boundaries converter wired for pipeline integration. 18.5 GiB/day average disk consumption; ext4 5% reserve enforcement pending sudo-required owner decision.

    featurehealth-check