Are Magnitude-Only Spectrograms Sufficient for Acoustic Classification in PAM? A Phase Retrieval Perspective

The passive acoustic monitoring (PAM) community has achieved remarkable success using magnitude-only short-time-Fourier-transform (STFT) spectrograms for underwater acoustic classification, raising the question of whether phase information provides discriminative value for this task. We apply recent phase retrieval theory to investigate this question, examining whether common underwater acoustic sources and PAM preprocessing parameters satisfy the mathematical conditions for implicit phase recovery from magnitude spectrograms. Our analysis reveals that a typical PAM configuration partially violates theoretical requirements for guaranteed phase retrieval, likely because practitioners optimize STFT parameters for frequency discrimination or visual interpretation rather than phase recovery. Despite these violations, magnitude-only approaches remain highly effective. However, whether this effectiveness stems from partial phase recovery or from phase information lacking discriminative value remains to be determined. Both explanations have important implications for parameter selection: if phase information is valuable but partially recoverable, optimizing STFT parameters for phase retrieval could improve performance; if phase lacks discriminative value, current magnitudeonly approaches may already be optimal. This work illuminates recent phase retrieval theory for investigating these competing hypotheses.