Audio Retrieval Using Spectrogram Energy Extreme Points

The patent segments the spectrogram into multiple local regions around each extreme point, extracting features from each region separately. This segmentation allows the system to capture local structural information while maintaining the speed benefits of extreme point-based methods, thereby improving matching accuracy without sacrificing retrieval speed.

The patent applies local quality by extracting features from specific local regions surrounding extreme points rather than using global features. This local feature extraction makes the retrieval more robust to noise while maintaining computational efficiency, as it focuses on localized patterns that are less susceptible to random noise interference.

The patent merges multiple local features from different regions around extreme points into a comprehensive feature representation. This combining approach maintains relatively low computational complexity while improving match accuracy by aggregating information from multiple local areas, making the system more robust to noise without requiring excessive computational resources.

The patent transitions from one-dimensional extreme point coordinates to two-dimensional local region features by extracting features from spectral bins around each extreme point. This dimensional expansion provides more discriminative power for matching while maintaining computational efficiency through the use of hash tables and compact feature representations.