Attribution Structure of Initials, Finals and Tones in Mandarin Speaker Verification Models
-
Abstract
The phonological information relied upon by deep speaker verification models for producing verification scores remains an open question in Chinese information processing. This study applies LayerCAM and CorrRISE, two general-purpose explainability methods, to a ResNet34 speaker verification model trained on CN-Celeb2. Phoneme-level alignments are obtained on AISHELL-3 via the Montreal Forced Aligner (MFA), and saliency sequences are aggregated to initials, finals, tones, and toned finals through a phoneme class importance distribution (PCID) metric. Two findings are developed. First, initial and final categories are interleaved in attribution ranking, forming a three-tier pattern rather than a clean initial–final split. Second, tonal attribution differences among tones 1–4 persist after controlling for finals, but the highest- and lowest-attributed tones are inconsistent across finals, a pattern better understood within a toned-final framework. Spearman rank correlations reach 0.82 for initials and 0.66 for finals, supporting cross-method consistency. We further validate the findings with ECAPA-TDNN. Under CorrRISE, the cross-model Spearman rank correlations on initials and finals reach 1.00 and 0.94, indicating that the main findings are not limited to ResNet34.
-
-