| 摘要 |
<p>A voice processing technique, that may be implemented as an algorithm or logic or a method, that electronically determines if a digitized sound source contains a voice signal component at periodic instances in time. The present invention uses spectral estimation techniques combined with statistical analysis and discrete decision logic to determine a voice or no-voice decision. The present invention uses a measurement of standard deviation of each energy band combined with a neural network to determine which bands may be used to detect voice modulation. Second-order statistics are used to determine which energy bands are most reliable for detection of a voice signal. The neural network is used to make the decision based on past values of these statistics. In addition, the voice decision logic controls the gain of its microphone input to provide a superior automatic gain control microphone. The present invention updates the AGC gain only during voice modulation. The present invention operates in noisy environments while maintaining high accuracy, resulting in very few false activations. The present invention adapts to the ambient noise level automatically, eliminating the need for a manual level control. The present invention makes decisions quickly, reducing problems associated with chopping off the beginning of words. The present invention controls the AGC function, and this control eliminates the effects of runaway gain when no voice signal is present, saturation when the circuit is reactivated, and also ensures consistent gain with speech containing long pauses between phrases. <IMAGE></p> |
