Sound retrieval and ranking using sparse auditory representations

Area of Science:

• Acoustics and Signal Processing
• Machine Learning for Audio Analysis
• Computational Auditory Scene Analysis

Background:

• Effective sound representation is crucial for systems understanding human auditory environments.
• Evaluating sound representations requires large-scale, quantitative frameworks.
• Machine vision techniques can be adapted for audio processing tasks.

Purpose of the Study:

• To quantitatively evaluate different sound representations in a large-scale sound-ranking task.
• To compare novel auditory models against conventional Mel-frequency cepstral coefficients (MFCCs).
• To adapt and apply the passive-aggressive model for image retrieval (PAMIR) to audio feature extraction.

Main Methods:

• Utilized a sound-ranking framework adapted from the passive-aggressive model for image retrieval (PAMIR).
• Compared adaptive pole-zero filter cascade (PZFC) auditory filter banks with sparse-code feature extraction.
• Evaluated stabilized auditory images with multiple vector quantizers against conventional MFCC front ends.

Main Results:

• Auditory models demonstrated a significant advantage over vector-quantized MFCCs.
• The best auditory model achieved 73% precision at top-1 and 35% average precision.
• This represents an 18% improvement compared to the best-performing MFCC front end.

Conclusions:

• Advanced auditory models offer superior performance for large-scale sound recognition tasks.
• Sparse-code feature extraction from auditory images provides a more discriminative representation than MFCCs.
• The PAMIR framework is effective for evaluating and developing robust auditory feature representations.