Skip to content

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.

← All 13 tasks · Quickstart