Audio Spectrum Classifier for Harmonic vs Noisy Mode Selection

The patent segments the music signal classification task into two distinct sub-tasks: (1) detecting peakyness using crest factor analysis to identify harmonic content, and (2) detecting noisiness using spectral flatness and zero-crossing rate to identify noise-like content. This segmentation allows each sub-task to be optimized independently, achieving fine-grained discrimination between harmonic and noise-like music segments that generic speech-music classifiers cannot accomplish.

The patent applies local quality analysis by examining specific spectral characteristics at different frequency bands and time moments. Instead of using a single global classifier, it analyzes local spectral flatness, zero-crossing rates, and crest factors at specific points in the spectrum to identify the local nature of the signal (harmonic vs. noise-like), enabling precise discrimination based on local signal properties.

The patent implements dynamic mode selection where the encoder adapts its behavior based on real-time signal characteristics. The mode decision logic dynamically switches between different encoding modes (e.g., harmonic mode for peaky signals, noise-like mode for non-peaky signals) based on the calculated crest factor and spectral flatness, allowing the codec to optimize performance for each signal type without requiring a completely complex static structure.

The patent changes encoding parameters based on signal characteristics by using the calculated peakyness and noisiness parameters as control signals. When crest factor exceeds a threshold (indicating peaky/harmonic signal), the system switches to a encoding mode optimized for such signals. When spectral flatness indicates noise-like characteristics, a different encoding mode is activated. This parameter-based adaptation optimizes encoding performance while maintaining manageable system complexity.

The patent replaces complex mechanical signal processing systems with more efficient mathematical operations in the frequency domain. Instead of using complex time-domain analysis or multiple sequential filters, it substitutes direct computation of spectral flatness and zero-crossing rate from frequency-domain coefficients, which are more computationally efficient and provide the necessary discrimination accuracy with reduced complexity.

The patent changes the approach by computing only the most critical spectral parameters (spectral flatness, zero-crossing rate, crest factor) rather than performing complete spectral analysis. This selective parameter extraction maintains high classification accuracy while significantly reducing computational burden, as these three parameters capture the essential characteristics needed to distinguish harmonic from noise-like signals.