Tripartite Graph Clustering for Audio Feature Extraction via Distributed Cache Networks

Zhenyue Qin

Abstract:

We propose a new method for extracting audio features using tripartite graph clustering on distributed cache systems. Current audio processing methods are slow when working with large datasets, especially for real-time applications that need fast results.
Our approach uses a three-part division that looks at spectral properties, time patterns, and semantic connections in audio signals at the same time. We build graphs with three types of nodes: audio segments, extracted features, and metadata. The connections between nodes are weighted based on how similar they are in both time and frequency domains.

We develop a distributed clustering algorithm that works across multiple cache layers. This allows parallel processing while keeping related data close together and reducing network traffic. The cache system stores commonly used feature patterns for quick access and computes rare combinations when needed.

We tested our method on standard audio datasets. Results show our approach is 3.2 times faster than traditional centralized methods while maintaining similar accuracy for tasks like speech recognition and music classification. The distributed design provides fault tolerance and can scale easily, making it good for edge computing and real-time audio processing. Our method also uses 40% less peak memory through smart cache management and feature sharing across parallel nodes.