Decision-threshold adaptation on dysarthric speech: evaluation code, verification instruments and metrics for a pre-registered study on TORGO

Abstract

Companion deposit for a pre-registered evaluation of decision-layer interventions in speaker verification for speakers with dysarthria, on the TORGO corpus. Contains: the frozen protocol with its complete dated changelog; the evaluation, scoring and metrics code; the LibriSpeech control run; the figure script; the verification instruments used to audit the work; and the metrics files the manuscript reports. Does not contain audio or any file derived from TORGO or LibriSpeech. Neither corpus is redistributed here. Manifests, trial lists, per-trial scores and speaker embeddings are regenerated by the released code from a local installation of the corpora; the SHA-256 digest of each is published in the README so that a regeneration can be verified against the run reported in the manuscript. See README.md for what runs from this archive alone, what requires a corpus, and which reported value traces to which file. MANIFEST.tsv lists every file with its size and SHA-256; verify_manifest.py recomputes them.