PHALAR: Phasors for Learned Musical Audio Representations

PHALAR's actual novelty lies in treating phase as a learnable invariance rather than discarding it as prior work does. Most audio models flatten phase information into magnitude spectrograms; this work encodes phase relationships directly into the contrastive loss, which is a concrete architectural choice, not just a marginal efficiency win.

This connects to a broader pattern in recent representation learning: moving from generic feature extraction toward domain-specific inductive biases. The Transformers paper from May 5th tackled selective early-representation access by making routing learnable rather than static, and the LASE work from May 1st addressed cross-script speaker identity by adding adversarial projection layers. PHALAR follows the same logic: instead of hoping a generic self-supervised objective captures musical structure, it bakes music-specific invariances into the embedding space itself. The downstream validation through beat tracking and chord probing mirrors how the encoding probe paper tests what models actually encode, rather than assuming learned features align with human intuition.

If PHALAR's gains replicate on held-out datasets from MoisesDB or Slakh that weren't used during development, and if downstream tasks like source separation or music tagging show consistent improvements without retraining, the phase-aware pooling primitive is genuinely reusable. If performance gains collapse on out-of-distribution music (e.g., non-Western instruments or atypical production), the inductive bias is too narrow for production systems.