Cite the corpus paper (arXiv:2608.01560) at the emotion-regression note it explains

README.md

@@ -148,7 +148,10 @@ only, with no audio-image supervision.
 - **Emotion recognition regressed** relative to v0.2. On RAVDESS zero-shot emotion the
   score is 0.21 (v0.2: 0.35). The v0.3 training mix improved spoken-content
   understanding at the cost of vocal-prosody sensitivity. If speech-emotion is central
-  to your use, evaluate before adopting.
+  to your use, evaluate before adopting. Why this happens, and what does and does not
+  recover it, is the subject of a dedicated paper: the corpus's discriminative
+  structure, not its size or caption vocabulary, decides which attributes the
+  embedding encodes ([arXiv:2608.01560](https://arxiv.org/abs/2608.01560)).
 - One evaluation cell, `CommonLanguageAgeDetection`, derives from Common Voice, which is
   part of the training data. It scores low (0.171) and does not inflate the reported
   mean, but the potential overlap is noted for completeness.
@@ -277,3 +280,15 @@ Technical report:
   year    = {2026}
 }
 ```
+
+Corpus analysis (what the audio training corpus does and does not teach):
+
+```bibtex
+@article{tonmoy2026axis,
+  title   = {Discriminative Axis, Not Data Volume: What a Contrastive
+             Corpus Teaches an Audio Embedding},
+  author  = {Tonmoy, Abdul Basit},
+  journal = {arXiv preprint arXiv:2608.01560},
+  year    = {2026}
+}
+```