?url_ver=Z39.88-2004&rft_val_fmt=info%3Aofi%2Ffmt%3Akev%3Amtx%3Adc&rft.title=Effective+Low-Cost+Time-Domain+Audio+Separation+Using+Globally+Attentive+Locally+Recurrent+Networks&rft.creator=Lam%2C+MWY&rft.creator=Wang%2C+J&rft.creator=Su%2C+D&rft.creator=Yu%2C+D&rft.description=Recent+research+on+the+time-domain+audio+separation+networks+(TasNets)+has+brought+great+success+to+speech+separation.+Nevertheless%2C+conventional+TasNets+struggle+to+satisfy+the+memory+and+latency+constraints+in+industrial+applications.+In+this+regard%2C+we+design+a+low-cost+high-performance+architecture%2C+namely%2C+globally+attentive+locally+recurrent+(GALR)+network.+Alike+the+dual-path+RNN+(DPRNN)%2C+we+first+split+a+feature+sequence+into+2D+segments+and+then+process+the+sequence+along+both+the+intra-+and+inter-segment+dimensions.+Our+main+innovation+lies+in+that%2C+on+top+of+features+recurrently+processed+along+the+inter-segment+dimensions%2C+GALR+applies+a+self-attention+mechanism+to+the+sequence+along+the+inter-segment+dimension%2C+which+aggregates+context-aware+information+and+also+enables+parallelization.+Our+experiments+suggest+that+GALR+is+a+notably+more+effective+network+than+the+prior+work.+On+one+hand%2C+with+only+1.5M+parameters%2C+it+has+achieved+comparable+separation+performance+at+a+much+lower+cost+with+36.1%25+less+runtime+memory+and+49.4%25+fewer+computational+operations%2C+relative+to+the+DPRNN.+On+the+other+hand%2C+in+a+comparable+model+size+with+DPRNN%2C+GALR+has+consistently+outperformed+DPRNN+in+three+datasets%2C+in+particular%2C+with+a+substantial+margin+of+2.4dB+absolute+improvement+of+SI-SNRi+in+the+benchmark+WSJ0-2mix+task.&rft.subject=speech+separation%2C+TasNet%2C+low-cost%2C+multi-head%0D%0Aattention&rft.publisher=IEEE&rft.date=2021&rft.type=Proceedings+paper&rft.language=eng&rft.source=+++++In%3A++Proceedings+of+the+2021+IEEE+Spoken+Language+Technology+Workshop+(SLT).++(pp.+pp.+801-808).++IEEE+(2021)+++++&rft.format=text&rft.identifier=https%3A%2F%2Fdiscovery.ucl.ac.uk%2Fid%2Feprint%2F10154105%2F1%2F2101.05014.pdf&rft.identifier=https%3A%2F%2Fdiscovery.ucl.ac.uk%2Fid%2Feprint%2F10154105%2F&rft.rights=open