first then collect all clean speech attributed to speaker zero zero in that recording ideally

excluding overlaps very short noises music low confidence regions etc now you may have three or

ten minutes of damoon's voice rather than a two second utterance create several embeddings from

substantial clean windows say 10 to 30 seconds and aggregate the result that gives you something

closer to open quote recording speaker zero damoon similarity 0.83 0.54 dad 0.49 close quote rather

than open quote 1.8 second utterance nobody above threshold close quote i strongly suspect this

alone would improve your identification rate enormously and don't store one embedding equals

one person for each known person i'd maintain a speaker profile containing many exemplars

for example damoon phone call clean speech embeddings office microphone embeddings dji

microphone embeddings car recordings quiet environment noisy environment then recognition

becomes something like nearest neighbor voting against perhaps 20 to 100 known exemplars

that matters because a voice embedding made from an iphone call through a codec can be surprisingly