Magor Communications Corporation

Kathryn Adeney

Dean Swan

An active stream is selected from one of a plurality of audio streams generated in a common acoustic environment by obtaining, for each stream obtaining, at a series of measurement instants t, where n=1 . . . N, a final performance metric that is representative of the stream's goodness for representing near-end speech at each measurement instant t. The final performance metrics for each stream are accumulated to determine an overall performance score. A switch to a new stream as the active stream occurs when the stream with the best overall performance score exceeds the overall performance score of the currently active stream by a threshold amount.

Method and apparatus for selecting an audio stream

Processing requirements for acoustic echo cancellation in voice communications are significant and are even more so as the bandwidth of the communication increases. Whilst voice communication occupies a relatively narrow band of frequencies the processing requirements and so forth for wideband communication render acoustic echo cancellation difficult to achieve in a cost effective manner. The invention provides for acoustic echo cancellation within wideband communications by dividing the communications into sub-bands and applying acoustic echo cancellation to some sub-bands whilst processing other sub-bands according to the status of the communications. Additional sub-bands are transmitted at either full-duplex or half-duplex.

Wideband speakerphone

Within audio telecommunications for audio-visual conferencing a means of reliable double-talk detection is required for several reasons, including control of residual echo, and control of adaptive algorithms used to estimate the echo. Existing techniques tend to either fail when echo return loss is very low or require highly complex algorithms to increase the range of detecting whether double-talk is occurring. The invention is based upon a new method of estimating the power envelope of the double-talk signal. This estimation is then employed in conjunction with a slightly modified version of an echo return loss enhancement (ERLE) based statistic, to form a double-talk decision based on a combination of the two statistics. The resulting decision rule works for both low-ERLE and high-ERLE situations, which is useful when there can be large variations in the magnitude of the echo path seen by the algorithms, or where near-end noise prevents the achievement of high ERLE.

Growing community of inventors

Dunrobin, Canada

Kathryn Adeney