T6A — Speech / ASR — mic & room shift¶
Paper evidence: main.pdf · Block findings
Lemma: D6 · Stack: pytorch
Nuisance key: temporal
Production change: Microphone, room, codec; transcript / word label fixed.
Notebook (Run All, built-in demo): t06a-speech-whisper.ipynb
pip install matching-pmh torch
# Open the notebook and Run All
What this task achieved (headline)¶
pmh_content_residual: Libri other-WER 14.63% (−8.6 pp vs baseline 23.26%); TDI 0.381.
| baseline WER | pmh_content_residual |
|---|---|
| 23.26% | 14.63% |
Note: Paper uses content-residual W (pmh.calibrate.content_residual_subspace). Demo uses temporal on sequence embeddings.
Subtasks (paper)¶
LibriSpeech four arms + WER¶
pmh_content_residual WER 14.63%.
Content-residual subspace¶
Only matched W fixes geometry.
Strengthening analysis JSON¶
Run with matching-pmh¶
from pmh import PMHTrainer, evaluate_robust_fit
# nuisance="temporal"
Do not use PMH when¶
Language or vocabulary change at deploy.
Replace demo data with yours¶
Swap demo loaders for your train_loader, source_batches, target_batches, and deploy holdout. Hook the backbone before your task head.