Ab­stract

Audio pro­cess­ing had been in de­mand through­out the elec­tronic era. Re­cent ad­vances in neural net­works in­creased the de­mand on audio pro­cess­ing for speech recog­ni­tion ap­pli­ca­tions. In this work, a rig­or­ous study on the dy­namic range and sys­tem com­plex­ity op­ti­miza­tion is pre­sented for a mixed-sig­nal key­word spot­ting sys­tem. The pro­posed sys­tem con­sists of an ana­log fea­ture ex­trac­tor and a neural net­work based key­word classifier. The re­sults showed that with the pro­posed method, more than an order of mag­ni­tude power sav­ing can be achieved in the ana­log fea­ture ex­trac­tion com­pared to the dig­i­tal state-of-the-art coun­ter­part.