Voice Activity Detection Based on Subband Spectral Entropy

During my postgraduate research, I designed and implemented a voice activity detection system based on subband spectral entropy analysis. The algorithm operates by first decomposing the audio signal into multiple frequency subbands using filter banks or wavelet transforms, then calculating the spectral entropy for each subband to measure signal randomness. By monitoring entropy thresholds across subbands, the program achieves robust endpoint detection with core functions including signal framing, FFT processing, entropy computation, and adaptive thresholding. This implementation demonstrates high accuracy in pinpointing speech boundaries while maintaining computational efficiency and stability across diverse acoustic environments. The system's modular architecture allows for easy integration with various speech processing applications. I take great pride in developing this robust solution and believe it contributes valuable methodology to the speech signal processing domain.