PM-ALIGN: Phase–magnitude embedding alignment for auto-tune detection

Accurate Auto-Tune detection in music recordings remains challenging because pitch correction can be subtle, content-dependent, and often obscured by accompaniment, timbral variation, and production effects. Most existing approaches rely primarily on magnitude-based time–frequency representations, such as log-mel spectrograms, and therefore make limited use of phase information related to pitch dynamics and instantaneous-frequency behavior. To address this limitation, we propose PM-ALIGN, a phase–magnitude embedding alignment framework for Auto-Tune detection. PM-ALIGN employs two modality-specific encoders: a magnitude branch operating on log-mel spectrograms and a phase branch operating on Dynamic Relative Phase (DRP) features derived from Short-Time Fourier Transform (STFT) phase. The framework learns a shared latent space by encouraging cross-modal agreement while preserving class-discriminative structure through semi-hard triplet loss. The main contributions are threefold. First, DRP is formulated as an image-like phase representation compatible with ImageNet-pretrained Convolutional Neural Network (CNN) backbones. Second, a Mel+DRP concatenation baseline is established to examine the complementary role of phase information beyond magnitude-only inputs. Third, an explicit–implicit alignment strategy is introduced to reduce phase–magnitude mismatch while maintaining class separability in the learned embedding space. During inference, fused modality embeddings are classified by nearest-prototype matching. Experiments on the Multi-lingual Dataset for Auto-Tune Detection in Music Recordings (ATDD) benchmark show that PM-ALIGN outperforms unimodal baselines and conventional fusion strategies under the evaluated setting. The results indicate improved segment-level Auto-Tune detection performance under the evaluated ATDD setting, with the highest recall, F1-score, and accuracy among the compared models.