Audio Fingerprinting via DCT Sign Correlation for Pitch Shift Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing audio identification systems are not robust enough to accurately identify distorted audio signals due to pitch shifting and tempo variations, as they rely on spectral content analysis which is sensitive to these types of distortions, leading to low matching rates and increased computational load.
Innovation Solution
The system generates probe audio fingerprints using a Short-Time Fourier Transform, filters them with a band-pass filter, applies a two-dimensional discrete cosine transform (DCT), and calculates DCT sign-only correlation to detect pitch shifting, allowing for robust identification of distorted audio signals by comparing the sign information of the probe fingerprint with reference fingerprints.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If spectral content analysis is used for audio identification, then the system can identify audio signals, but it is not robust to pitch shifting and tempo variations
Solution Approach 1:
The patent transforms the audio fingerprint representation from spectral domain to time-domain envelope features. Specifically, it extracts amplitude envelopes from short-time Fourier transform magnitudes, then applies discrete cosine transform to convert to a different parameter space that is invariant to pitch and tempo changes. This parameter transformation makes the identification robust to distortions while maintaining accuracy.
2Measurement precision
If multiple fingerprint indexes are extracted for different pitch shifts, then matching rate improves, but computational load and storage space increase
Solution Approach 1:
The patent extracts only the essential time-domain envelope information from the spectral content, discarding the frequency-specific details that cause sensitivity to pitch shifts. By taking out only the amplitude envelope trajectory and representing it through DCT coefficients, the system achieves pitch-shift invariance without needing to store or process multiple pitch-specific fingerprint versions.
Solution Approach 2:
The time-domain envelope representation serves as a universal fingerprint that works for both original and pitch-shifted versions of audio signals. A single envelope-based fingerprint can match multiple pitch variations of the same audio content, making the system multi-functional without requiring separate processing paths for different pitch conditions.
Data Source
AI summary
An audio identification system generates a probe audio fingerprint of an audio signal and determines amount of pitch shifting in the audio signal based on analysis of correlation between the probe audio fingerprint and a reference audio fingerprint. The audio identification system applies a time-to-frequency domain transform to frames of the audio signal and filters the transformed frames. The audio identification system applies a two-dimensional discrete cosine transform (DCT) to the filtered frames and generates the probe audio fingerprint from a selected number of DCT coefficients. The audio identification system calculates a DCT sign-only correlation between the probe audio fingerprint and the reference audio fingerprint, and the DCT sign-only correlation closely approximates the similarity between the audio characteristics of the probe audio fingerprint and those of the reference audio fingerprint. Based on the correlation analysis, the audio identification system determines the amount of pitch shifting in the audio signal.