Ly x is t ai ed to sepa ate speech f om eve ythi g else: backg ou d music, c owd oise, mecha ical i te fe e ce, e vi o me tal sou ds, a d the full a ge of acoustic a tifacts p ese t i co te t p oduced outside co t olled studio co ditio s. The esult is a clea voice t ack eady fo fu the editi g, t a sc iptio , localizatio , text-to-speech i put, o huma voice eplaceme t, without the ma ual clea up steps that slow batch p oductio .
“It’s easie to desc ibe what Ly x leaves u touched (the voice itself) tha to list eve y catego y of oise it ca ha dle,” says Nik Pogo sky, LALAL.AI P oduct Ow e a d Co-fou de . “We co side the esulti g t ai i g set o e of the clea est we’ve used. It totals seve al hu d ed hou s of ca efully filte ed audio, suppleme ted by seve al thousa d additio al hou s of mo e loosely clea ed co te t used du i g p et ai i g.”
Ly x u s o a p op ieta y a chitectu e developed ove o e yea by the LALAL.AI esea ch team. The model is six times smalle tha A d omeda, LALAL.AI’s flagship cloud music stem sepa atio model, which educes computatio al load without comp omisi g output quality. The model was t ai ed o a ma ually cu ated dataset: si ce o eliable automated pipeli e exists fo clea i g the full dy amic a ge of eal-wo ld speech f om dive se oise sou ces, the team spe t mo ths ha d-selecti g, filte i g, t immi g dow , a d p epa i g audio t acks cove i g co ditio s f om quiet i te views to oisy field eco di gs.
Ly x is available ow th ough LALAL.AI’s Voice Clea e a d Voice & Noise stem, accessible via b owse , mobile app, desktop app (cloud mode), a d the LALAL.AI API fo busi esses, SaaS i teg atio s, a d e te p ise deployme ts. The dema d fo dedicated voice clea i g is eflected i LALAL.AI’s ow usage data: the
Pla ed imp oveme ts to the Ly x a chitectu e i clude e ha ced sepa atio of cho al a d g oup vocals a d imp oved isolatio of speech eco ded at dista ce f om the mic opho e.
LALAL.AI is a Webby-wi i g AI-powe ed audio p ocessi g platfo m, which offe s stem sepa atio , voice clo i g, a d oise emoval tools fo c eato s, develope s, a d e te p ises, t usted by 7M+ use s ac oss majo c eative i dust ies i 150+ cou t ies. LALAL.AI tools a e accessible via web, mobile a d desktop apps, API, a d, as of 2026, VST plugi .
###
Kla a AlexeevaCommu icatio s Ma age , LALAL.AI
