Rethinking Automated Voice Similarity by Shifting from EER to Embedding Geometry
Voice-similarity alignment tracks embedding dimensionality, not verification EER; a bottleneck lifts it from 0.08 to 0.74.
The paper argues that speaker-verification models are poor automatic proxies for human voice similarity even when verification accuracy is high. Perceptual alignment is driven by the learning objective rather than equal error rate: standard margin-based classification losses such as AAM-Softmax align much worse than prototypical metric losses. Effective dimensionality tracks human alignment with a -0.95 rank correlation, and imposing a dimensionality bottleneck lifts alignment from 0.08 to 0.74.
- Perceptual alignment depends more on the training objective than on equal error rate.
- Margin-based losses such as AAM-Softmax align worse with human voice judgments than prototypical metric losses.
- Effective embedding dimensionality correlates with perceptual alignment at rank correlation -0.95.
- A dimensionality bottleneck raises alignment from 0.08 to 0.74.
From the original article
1 min read at huggingface.co
Speaker verification (SV) models are commonly assumed to better capture nuances among speaker characteristics as verification accuracy improves, leading to their widespread use as automated proxies for human voice similarity in speech generation tasks.
By Szu-Chi Chen, Jia-Kai Dong, Yi-Cheng Lin · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.