An Optimized Hybrid Deep Feature Fusion Framework for Hardware-Efficient Music Retrieval
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
This paper introduces an optimized hybrid deep feature fusion framework designed for hardware-effi;cient music retrieval, specifically architected for heterogeneous AMD Xilinx FPGA platforms. To overcome the computational bottlenecks of probabilistic indexing in large-scale retrieval, the framework integrates three specialized hardware-accelerated engines, orchestrated via asynchronous ping-pong buffering and cycle-accurate AXI4-Stream scheduling to seamlessly synchronize dataflow across the PL and AIE domains: (1) a pipelined hybrid feature fusion module synthesized via Vivado HLS, which dynamically merges deep convolutional embeddings with statistical spectral descriptors to generate robust acoustic signatures, achieving a 93.8% retrieval precision with a processing latency of only 2.1 ms for 30-second audio segments; (2) an AI Engine-accelerated Variational Inference Processor that executes complex logistic normal approximations required by Correlated Topic Models (CTM) over an AXI4-Stream interconnect, optimizing covariance matrix operations to reduceBRAMutilization by 38% compared to standard architectures while sustaining a throughput of 480 audio segments per second; and (3) a DSP-optimized Topic Likelihood Estimator utilizing DSP48E2 slices to parallelize posterior probability calculations for rapid similarity matching, reducing inference latency by 5:6×relative to software-based Gibbs sampling and capping total power consumption at 14W. Validated on three benchmark music datasets (GTZAN, FMA-Small, and MagnaTagATune), the framework demonstrates a 91.5% Top-1 retrieval accuracy, outperforming baseline CNN-only approaches by 12.7% and achieving a 45% reduction in prediction variance across diverse musical textures. The system delivers 431× and 20× higher energy efficiency (29 mJ per inference) than equivalent high-performance CPU- and GPU-based solutions, respectively, along with a 48× smaller memory footprint (8.1 MB), facilitating deployment in edge-based music discovery systems. The complete hardware-software co-design has been verified on Xilinx Zynq UltraScale+ MPSoC devices using the Vitis 2022.2 toolchain.
Fields of Study & Subjects
Music and Audio Processing
Embedded Systems Design Techniques
Embedded Systems and FPGA Applications
Publication details
- Journal / source
- IEEE Access
- Volume
- 14
- Pages
- 52427–52439
- Publication date
- 1 January 2026
- Language
- EN