Nnabata Ọkà Okwu ECAPA-TDNN
ECAPA-TDNN bụ ụlọ ọrụ netwọkụ akwara na-atụgharị obere okwu ọ bụla ka ọ bụrụ ntinye 'mkpụrụokwu' kọmpat, na-enyere igwe aka ịkọ onye na-ekwu okwu.
Nchịkọta
It set the state of the art for speaker verification and remains the workhorse behind voice ID systems today.
Ime miri emi
ECAPA-TDNN na-anọchi anya nlebara anya ọwa mesiri ike, Mgbasa na mkpokọta na netwọkụ oge na-egbu oge, nke Desplanques na ndị ọrụ ibe webatara na 2020. Ọ na-ewuli na usoro x-vector ochie mana ọ na-agbakwunye nkwalite igodo atọ: Squeeze-Excitation blocks that reweight feature channels, multi-layer, shallow feature na aggre Nchịkọta nlebara anya n'ọwa na-adabere na ọnọdụ nke na-achịkọta okwu ogologo agbanwere n'ime otu vector kwụpụrụ. A zụrụ ya site na mfu ihe mgbakwunye-margin softmax (AAM-softmax) na nnukwu ụlọ ọrụ dị ka VoxCeleb, ọ na-emepụta ntinye ebe otu mkpachị igwe okwu na-agbakọta nke ọma. Atụnyere ụda olu abụọ na myirịta cosine. Na ule VoxCeleb1, ọ kpalitere ọnụego njehie nha anya n'okpuru ihe dị ka pasenti 1, nnukwu mmụba n'elu sistemu ndị gara aga.
Nghọta nka nka
Isi aghụghọ bụ ịgbakọ ọnụ ọgụgụ nlebara anya: kama ịgbakọ atụmatụ ọkwa etiti, netwọkụ ahụ na-amụta nha nlebara anya nke ọ bụla nke mere ka okpokolo agba dị mkpa (okwu a na-ekwu nke ọma) gụta karịa ịgbachi nkịtị ma ọ bụ mkpọtụ, mgbe ahụ ọ na-agbakọ ma ọ̀tụ̀tụ̀ dị arọ na ngbanwe ọkọlọtọ. Ihe mgbochi SE na ụdị Res2Net nwere ọtụtụ mgbanwe na-eme ka oyi akwa ọ bụla nwee ọnọdụ n'ọnọdụ nkwupụta okwu zuru ụwa ọnụ. Ntinye ikpeazụ na-abụkarị akụkụ 192, nke ejiri anya cosine gbara ya.
Mmetụta atụmatụ
Nweta na iru
Ọ na-eme ka nnweta ya dịkwuo mma site na ndegharị, ịkọ akụkọ, na ntụgharị olu.
Ọnụ ego na mmefu ego
Ndị otu mgbasa ozi nwere ike ibubata ọdịyo a na-egbu maramara ngwa ngwa site na iji obere mmefu ego.
Ọsọ na ọnụ ọgụgụ
Sistemụ na-eche ihu ndị ahịa nwere ike hazie mkparịta ụka n'ọtụtụ buru ibu.
Ọdịnihu nke nnabata Ọkà Okwu ECAPA-TDNN
Nchọpụta na-aga n'ihu n'ihu-nlekọta onwe ya dị ka WavLM na wav2vec 2.0 na-enye azụ azụ azụ ECAPA ụdị, nke na-egbutu data akpọrọ achọrọ ma na-abawanye ike na mkpọtụ na obere vidiyo. Na-atụ anya mwekota siri ike na mgbochi mkpachapụ ka otu ihe nlereanya na-achọpụta ma gosipụta onye na-ekwu okwu, ụdị dị ntakịrị maka iji ngwaọrụ, yana ịrụ ọrụ ziri ezi siri ike iji belata oghere njehie n'ofe ụda olu, afọ, na asụsụ ka ụda biometrics na-agbasawanye n'ime ụlọ akụ na njikwa ohere.
Mmejuputa n'ezie n'ụwa
Nbanye biometric olu maka ụlọ akụ ekwentị, ebe ụda olu onye na-akpọ oku dabara na ndebiri edebanyere aha kama ịbụ PIN.
Ntugharị okwu n'ime ngwaọrụ ntụgharị okwu nzuko, na-akpọ 'onye kwuru mgbe' site na ịchịkọta ECAPA ntinye.
Nnwale nyocha n'ihu ọha na nke etiti oku iji gosipụta ma ndekọ abụọ sitere n'otu onye.
Ịkwanye ntụziaka nkwenye nke ọkà okwu na ngwa ngwa mepere emepe dị ka SpeechBrain na Kaldi maka ndị nyocha na ndị mmalite.
Ihe ize ndụ & okporo ụzọ nche
Iji olu eme ihe na ihe egwu mpụta ga-abawanye mgbe nkwenye na-efu.
Izi ezi nwere ike ịdaba n'ofe ụda olu, olumba ma ọ bụ gburugburu mkpọtụ.
Enwere ike imehie ọdịyo sịntetik dị ka ezigbo okwu na-enweghị akara doro anya.
Map mmejuputa
Nweta nkwenye doro anya maka ijide olu, imechi, na ijigharị.
Nwale ogo n'ofe ndị na-ekwu okwu dị iche iche yana ọnọdụ ndabere.
Kọwaa mgbe mmadụ ga-enyocha ma ọ bụ kwado nsonye.
Deba aha ọdịyo sịntetik ma debe ndekọ ihe ndekọ maka ịza ajụjụ.
Nọgide na-eme nchọpụta
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the ECAPA-TDNN Speaker Recognition quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
Ntuziaka na-esote
Nkwekọrịta ụda olu
Ajụjụ a na-ajụkarị
What is ECAPA-TDNN Speaker Recognition?
ECAPA-TDNN bụ ụlọ ọrụ netwọkụ akwara na-atụgharị obere okwu ọ bụla ka ọ bụrụ ntinye 'mkpụrụokwu' kọmpat, na-enyere igwe aka ịkọ onye na-ekwu okwu. O debere ọnọdụ nka maka nkwenye ọkà okwu ma bụrụkwa ihe na-arụ ọrụ n'azụ sistemụ NJ olu taa.
Kedu ihe bụ isi mmalite nke ụdị ECAPA-TDNN maka obere okwu enyere?
ECAPA-TDNN na-esepụta okwu ogologo ogologo n'ime otu vector etinyere ogologo ogologo nke na-anọchi anya njirimara olu ọkà okwu.
Kedu ka esi agbakwunyere ECAPA-TDNN abụọ ma e jiri ya tụnyere ikpebi ma ha bụ nke otu ọkà okwu?
Mgbe ewepụtara ihe mgbakwunye, myirịta cosine (ma ọ bụ akara metụtara ya dị ka PLDA) na-atụle ka mpụta olu abụọ ahụ si dị nso.
Kedu ihe oyi akwa 'attentive statistics pooling' na-eme?
Ịkọkọ ọnụ ọgụgụ nlebara anya na-ekenye nlebara anya nlebara anya na okpokolo agba wee chịkọta ha ka ọ bụrụ ihe dị arọ na ngbanwe ọkọlọtọ, na-emesi okpokolo agba ozi ike.
Kedu dataset nke a na-ejikarị azụ na benchmark ECAPA-TDNN ụdị ọkà okwu?
VoxCeleb, nnukwu vidiyo ajụjụ ọnụ nke ndị ama ama ewepụrụ na vidiyo, bụ ọkọlọtọ corpus maka ọzụzụ na nyocha sistemu nkwenye ọkà okwu.
Kedu ihe mgbochi 'SE' (Squeeze-Excitation) na-enye aka na nhazi ụlọ?
Squeeze-Excitation blocks na-amụta itule ọwa njirimara ọ bụla site na iji ozi zuru ụwa ọnụ, na-eme ka netwọk ahụ mesie ọwa kachasị bara uru.