Failed on forth step - validate

[brantje]

09:36:28  INFO      ============================================================
09:36:28  INFO      [3/13]  download  —  Download datasets, Piper TTS model, tools
09:36:28  INFO      ============================================================
09:36:28  INFO        Repo already cloned: /home/node/openwakeword-trainer/data/piper-sample-generator
09:36:28  INFO        Downloading Piper LibriTTS model (~800 MB) …
  Piper LibriTTS model (~800 MB): 195/195 MB (100%)
09:36:30  INFO        Saved /home/node/openwakeword-trainer/data/piper-sample-generator/models/en_US-libritts_r-medium.pt
09:36:30  INFO        Already downloaded: openwakeword_features_ACAV100M_2000_hrs_16bit.npy
09:36:30  INFO        Already downloaded: validation_set_features.npy
09:36:30  INFO        MIT RIRs already present
09:36:30  INFO        AudioSet subset already present
09:36:30  INFO        FMA subset already present
09:36:30  INFO      [3/13]  download  ✓ PASSED
09:36:30  INFO      
09:36:30  INFO      ============================================================
09:36:30  INFO      [4/13]  verify-data  —  Check all data files present & minimum sizes
09:36:30  INFO      ============================================================
09:36:30  INFO        OK: openwakeword_features_ACAV100M_2000_hrs_16bit.npy             16479 MB
09:36:30  INFO        OK: validation_set_features.npy                                   176 MB
09:36:30  ERROR       TOO SMALL: /home/node/openwakeword-trainer/data/piper-sample-generator/models/en_US-libritts_r-medium.pt  (204089915 bytes, expected ≥600000000)
09:36:30  ERROR       EMPTY dir: /home/node/openwakeword-trainer/data/mit_rirs
09:36:30  INFO        OK: audioset_16k/                                                 200 files
09:36:30  INFO        OK: fma_small/                                                    100 files
09:36:30  INFO        OK: piper-sample-generator installed
09:36:30  ERROR     [4/13]  verify-data  ✗ FAILED
09:36:30  ERROR     
09:36:30  ERROR     Pipeline stopped.  Fix the issue above, then resume:
09:36:30  ERROR       python train_wakeword.py --from verify-data

[newellnarco]

the trainer's MIN_SIZES threshold (600 MB, "~800 MB" comment) is wrong. mine was ~204MB.

[newellnarco]

A few pipeline bugs hit on a clean run (great tool otherwise — thanks!). Found four issues; fixes/suggestions for each:

1. verify-data rejects a valid Piper model. MIN_SIZES["piper-sample-generator/models/en_US-libritts_r-medium.pt"] is 600_000_000 (comment says "~800 MB"), but the actual v2.0.0 GitHub release asset is 204,089,915 bytes (~195 MiB) — I confirmed the Content-Length on github.com/rhasspy/piper-sample-generator/releases/download/v2.0.0/en_US-libritts_r-medium.pt. A complete, correct download fails as TOO SMALL. Suggest lowering the floor to ~150 MB.

2. Empty directories are skipped, never retried. In step_download, the checks for mit_rirs / audioset_16k / fma_small are if not

   .exists(): download() else "already present". A prior failed download that left an empty dir is then treated as "already present", and verify-data later hard-fails with "EMPTY dir" — no self-heal short of manual rm. Suggest if not d.exists() or not any(d.iterdir()).

3. _download_mit_rirs contradicts verify-data. On any exception it logs "Creating empty RIR directory — training will proceed without RIRs" and makes an empty dir — but step_verify_data treats an empty mit_rirs as a hard failure, so the "proceed without RIRs" path can never actually proceed. Also, load_dataset("davidscripka/MIT_environmental_impulse_responses", trust_remote_code=True) fails on current datasets (script-based/trust_remote_code datasets are restricted/removed), so this path triggers for everyone. Either make verify-data tolerant of an empty/absent mit_rirs (since the downloader already treats RIRs as optional) or update the RIR source.

4. generate step → ModuleNotFoundError: No module named 'generate_samples'. openwakeword.train imports generate_samples (which lives in piper-sample-generator), but oww_wrapper.py doesn't add data/piper-sample-generator to sys.path, and the current rhasspy/piper-sample-generator default branch no longer has generate_samples.py at its root (raw URL 404s). Suggest pinning piper-sample-generator to a commit that still ships generate_samples.py and prepending that dir to sys.path/PYTHONPATH before delegating to openwakeword.train.

Happy to send a PR for 1–3 if useful.

[karthikeyareddy-create]

@newellnarco i want to know what augmentations did you do and what were the db values for background noise did you set?