Rethinking Automated Voice Similarity by Shifting from EER to Embedding Geometry
The paper argues that voice-similarity systems can score well at verification while still missing what humans hear.
The authors say EER does not reliably track human judgments of voice similarity. Their analysis points instead to the geometry of speaker embeddings, especially effective dimensionality. Models trained with standard margin-based classification losses aligned worse with perception than prototypical metric-loss models. A dimensionality bottleneck raised their alignment measure from 0.08 to 0.74. HF Daily Papers' note
The authors say EER does not reliably track human judgments of voice similarity. Their analysis points instead to the geometry of speaker embeddings, especially effective dimensionality. Models trained with standard margin-based classification losses aligned worse with perception than prototypical metric-loss models. A dimensionality bottleneck raised their alignment measure from 0.08 to 0.74. HF Daily Papers' note
score 4