摘要 |
A method of detecting voice in an audio signal comprises the steps of determining an average peak value representing an envelope of the audio signal, determining a running instance of audio signal standard deviation, which corresponds to one of a number of overlapping time intervals, and updating a power density function (PDF) by adding instances of noise to the PDF if the average peak of the audio signal exceeds the current level of the audio signal by a certain amount and if the current standard deviation value fails below a threshold for a predetermined time interval. A noise floor is located based on the mean value of the PDF, and, if the audio signal sustains a power level exceeding the noise floor, voice activity is determined to be present in the audio signal. The PDF is updated by a low confidence factor if all of the standard deviation values calculated during a certain period of time are below the threshold value and by a high confidence factor if all standard deviation values within a certain longer period of time period are below the threshold value.
|