Cover Song Recognition: Temporal Slope Features and Delta-Gradient Optimization

Cover song recognition typically relies on computationally expensive raw audio analysis, which limits applicability in resource-constrained or privacy-preserving scenarios. Existing audio-free alternatives use metadata or lyrics to augment rather than replace audio analysis, and the performance level achievable from compact pre-computed audio descriptors alone has not been systematically established. This paper is an empirical study of that constrained regime: how far pre-computed audio descriptors go without raw audio, and which engineering choices matter. We propose a lightweight approach based on a Siamese retrieval model operating on Million Song Dataset features and evaluated on the SecondHandSongs benchmark. The method combines temporal slope features extracted from pitch and timbre time series, confidence-weighted feature multiplication, and a hybrid delta-gradient optimization framework designed for low-dimensional feature spaces. Each component contributes measurably: temporal slope features raise the unweighted baseline from 0.370 to 0.400 MAP@10, normalization to 0.420, heuristic feature weighting to 0.520, and delta-gradient refinement to the final 0.553, to our knowledge, the best result reported under this constraint. This remains below audio-based state-of-the-art systems, which exploit fine-grained spectral detail that compact descriptors discard; in exchange, the proposed system requires much less storage and computation, enabling privacy-preserving, bandwidth-constrained, and audio-unavailable deployment scenarios.