jr-public/jrich-wakeword-trainer
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases 4 Wiki Activity

4 Releases 4 Tags

RSS Feed
  • v0.2.1-jrich-5 3a568d150b
    Compare

    JRich Wakeword Trainer 0.2.1-jrich-5 Stable

    westtexasfish released this 2026-08-15 02:32:24 -05:00 | 0 commits to master since this release

    Resource-evidence correctness release. This supersedes 0.2.1-jrich-4.

    Fixes:

    • whole-device GPU memory is labeled as device-scoped; a pre-execution baseline and observed device deltas are recorded without claiming process attribution;
    • local run disk evidence is refreshed after the client and retraining bundles are written;
    • unmeasured local evaluation values are null with an explicit not_measured status instead of fabricated durations and zeros;
    • hardware guidance is now 8 GB VRAM, 8 CPU cores minimum / about 16 recommended for the measured completion time, 32 GB RAM, and 40 GiB free.

    Downloads:

    • Linux x86_64 / WSL2: jrich-wakeword-trainer-0.2.1-jrich-5.tar.gz
    • Windows with WSL2: jrich-wakeword-trainer-0.2.1-jrich-5.zip
    • trainer-release-v1.json records immutable runtime and archive hashes
    • SHA256SUMS verifies every release asset

    The launcher pulls the public runtime by immutable digest:
    git.jimandkrista.com/jr-public/jrich-wakeword-training@sha256:e5854a420060c897d192cd6bb3d85fba079a521587ebc1f627691058fa34fd8a

    The exact image and archives passed non-root/read-only, checksum, metadata, permissions, and archive-only import gates. User phrases, generated speech, checkpoints, and ONNX outputs remain local.

    Downloads