Voice extraction by on-line signal separation and recovery

摘要

The paper presents a formulation and an implementation of a systemnfor voice output extraction (VOX) in real-time and near-real-timenrealistic real-world applications. A key component includes voice-signalnseparation and recovery from a mixture in practical environments. Thensignal separation and extraction component includes several algorithmicnmodules with a variety of sophistication levels, which include dynamicnprocessing neural networks in tandem with (dynamic) adaptive methods.nThese adaptive methods make use of optimization theory subject to thendynamic network constraints to enable practical algorithms. Thenunderlying technology platforms used in the compiled VOX software cannsignificantly facilitate the embedding of speech recognition into manynenvironments. Two demonstrations are described: one is PC-based and isnnear-real-time, the second is digital signal processing based and isnreal time. Sample results are described to quantify the performance ofnthe overall systems