Voice Activity Detection Using Multi-Feature Voting

The voice detection system segments the audio signal processing into distinct feature extraction modules (zero-crossing detection, periodicity detection, energy ratio calculation, envelope-to-floor determination) and combines their outputs through voting. This segmentation allows each module to be simple and inexpensive while collectively achieving high detection accuracy, resolving the contradiction between accuracy and complexity.

The patent merges multiple simple detection algorithms into a unified voice activity detection system where each algorithm processes the audio signal independently and contributes to a final vote. This combining approach achieves accuracy comparable to complex algorithms while maintaining low implementation cost, as the individual algorithms are computationally inexpensive but their collective decision is robust.

The patent combines multiple simple algorithms (zero-crossing detection, periodicity detection, energy ratio calculation, envelope-to-floor determination) that individually have low computational complexity. By merging their outputs through a voting mechanism, the system achieves high detection accuracy without requiring any single complex algorithm, thus maintaining low device cost while improving accuracy.

The system divides the detection task into separate simple modules, each performing a basic operation on the audio signal. This segmentation allows the use of simple algorithms throughout while achieving high overall accuracy through the collective strength of multiple simple detectors, avoiding the need for expensive complex algorithms.

The patent segments the audio signal into frames and applies different simple feature extraction algorithms to each frame independently. This segmentation allows parallel processing and reduces the computational burden on any single algorithm, thereby reducing overall processing time while maintaining high accuracy through the combination of multiple features.

The system applies multiple feature extraction algorithms that may be slightly excessive for the task, but each algorithm is simple and fast. The voting mechanism then filters out redundant information. This partial/excessive action approach ensures high accuracy through multiple perspectives while keeping processing time low because each individual algorithm is computationally inexpensive.