VeriSpoof Uyasebenza

I-VeriSpoof iyimodeli enkulu yomsindo enomthombo ovulekile eyakhiwe yiNyuvesi yaseMelbourne, enamapharamitha angu-2.85B kodwa yaqeqeshwa ngamasampuli angu-29.7M kuphela. I-VeriSpoof ihlanganisa uphenyo lokukhohliswa kwezwi lube ukuphendula imibuzo yemodeli; izisindo ezifanayo zenza ukutholwa kobuqiniso, ukulandelela uhlelo, ukuqinisekisa umsuka ofanayo, ukuhlonza isiteshi kanye nokutholwa kokucabanga okunobufakazi obungaqinisekiswa. Ku-Speech-DF-Arena, i-EER ehlanganisiwe ingu-2.14%; ikleliswe endaweni yesithathu phakathi kwezinhlelo ezingu-28 futhi ingeyokuqala phakathi kwezinhlelo ezinomthombo ovulekile.

Ukutholwa kobuqiniso: inquma ukuthi ukuqoshwa kuyizwi lomuntu wangempela noma umsindo owenziwe yi-AI. Isikolo sivela emgqeni womehluko wokungenzeka kwe-log phakathi kwemisho yezimpendulo ezimbili ethi “okukhohlisiwe/umuntu wangempela”. Ku-Speech-DF-Arena, i-EER ehlanganisiwe ingu-2.14%, okuyindawo yokuqala phakathi kwezinhlelo ezinomthombo ovulekile.
Ukulandelela uhlelo lokukhiqiza: ihlonza ijeneretha ethile phakathi kwezinhlelo ze-TTS/VC ezingabhekwa, futhi ingahlukanisa ukuhlonza kube amazinga amathathu: imodeli ye-acoustic, i-vocoder kanye nohlelo. Ku-MLAAD, ukunemba kokuhlukanisa amajeneretha angu-12 abonwe ngaphambilini kungu-96.8%.
Ukuqinisekisa umsuka ofanayo: inquma ukuthi ukuqoshwa okubili kuvela ohlelweni olufanayo lokukhiqiza yini. Kuma-pair amajeneretha angazange abonwe ekuqeqeshweni, i-EER ingu-6.2%, ephansi ngo-77% kuneye-probe yama-pair eqeqeshwe kuzici ezimisiwe zomtshina ozinikele.
Ukuhlonza izimo zokuqoshwa: ihlonza umsindo ongemuva, i-reverberation nezinhlobo ze-codec kumsindo; ukunemba kwenguqulo ye-Plus kungu-81.3%.
Ukucabanga okuqinisekisekayo: ikhipha ukucabanga ngezinkambu ezimisiwe zokubuka, isiteshi, indawo, iphrofayela yomthombo nesiphetho. Umugqa ngamunye ungaqinisekiswa ngokulinganisa i-waveform noma ngamalebula edathasethi, futhi inikeza indawo yezingxenye ezikhohlisiwe njalo ngemizuzwana engu-0.16. Lokhu kukhuphula izinga lokutholwa kokukhohliswa kwasendaweni lisuka ku-78.4% laya ku-99.6%.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lokushintshisana ngamaphrojekthi e-AI anomthombo ovulekile
Ukwenza amathokheni omsindo ngesinqumo esikabili: i-encoder isebenzisa i-AuT encoder ye-Qwen3-ASR-1.7B. Izimo ezifihlekile zezingqimba ezingu-25 zihlanganiswa ngezisindo ezifundekayo bese zithunyelwa ku-audio tokenizer eqeqeshwe kusukela ekuqaleni. I-tokenizer ikhipha amathokheni omusho wonke angu-16 aphethe ulwazi lomthombo olungaguquki emhlabeni wonke, kanye namathokheni ozimele bezinhlaka angu-52, ngamunye omele imizuzwana engu-0.16 futhi ugcine ulwazi lwesikhathi. Umaka wesikhathi wombhalo ufakwa njalo ngemizuzwana engu-0.8 phakathi kwamafreyimu, ukuze imodeli yolimi ikwazi ukusho ngokuqondile indawo yezingxenye ezikhohlisiwe “ngemizuzwana”.
Impendulo uqobo iyisona sikolo: imodeli yolimi yi-MiniCPM5-2B ene-LoRA. Izinhloso zokuqeqesha kuyo yonke imisebenzi ukubikezela ithokheni elandelayo, ngaphandle kwanoma iyiphi i-classification head eqondene nomsebenzi. Isikolo sokutholwa siwumehluko wokungenzeka kwe-log phakathi kwemisho yezimpendulo zokukhohliswa nomuntu wangempela; isikolo somsuka ofanayo siwumehluko phakathi kuka-yes no-no; emibuzweni enezimpendulo eziningi, wonke umusho wempendulo ekhethiwe uyalinganiswa bese kuthathwa ophezulu.
Uhlelo lokuqeqesha lwezigaba ezintathu: esigabeni sokuqondanisa kokubona, imodeli yolimi iyamiswa, kuqeqeshwe kuphela i-tokenizer ne-DoRA ye-encoder, ukuze uhlangothi lomsindo luqale lufunde ukuphatha izimpawu zokukhohliswa. Kamuva kwengezwa ukuqeqeshwa kwe-LoRA yemodeli yolimi kuyo yonke imibuzo yophenyo, okuhlanganisa ijeneretha, imodeli ye-acoustic, i-vocoder nemibuzo yama-pair emazingeni amathathu. Ekugcineni, kwengezwa ukucabanga okunobufakazi namasampuli aphikisanayo ngesilinganiso sokufunda esiphansi, kuyilapho imibuzo yesigaba sesibili iphinde idlalwe ukuze ikhono lokuphendula ngesinyathelo esisodwa lingehli.
Inhloso yokucabanga ebalwa kusignali: umbhalo wokucabanga ubhalwa ngezinkambu ezimisiwe, futhi uhlobo ngalunye lwesitatimende lunomhloli ozimele ngemodeli. Umugqa wokubuka uqathaniswa nezilinganiso ze-bandwidth, i-SNR, izinga lesignali ne-clipping ku-waveform; umugqa wesiteshi uqathaniswa nezimo eziqoshwe ngesikhathi sokwandiswa kwedatha; umugqa wendawo uqathaniswa namalebula edathasethi; iphrofayela yomthombo iqathaniswa nemethadatha yejeneretha. Amasampuli aphikisanayo aguqula indawo ibe okuphambene noma afake ingxenye entsha, kudingeke ukuthi isiphetho siguquke ngokufanayo; uma kushintshwa izinkambu ezingahlobene kuphela, isiphetho kufanele singashintshi. Lokhu kuvinjelwa kwesignali yokuqeqesha kwenza isiphetho sincike kuphela ekutheni “kuphi okwenziwe ngokwenziwa”.

Ukufaka ikhodi: sebenzisa hf download AustinXiao/VeriSpoof-Plus --include "code/*" --local-dir verispoof ukuze ulande ikhodi, bese usebenzisa pip install ./verispoof/code ukufaka iphakheji yokuqagela.
Ukulayisha imodeli: ku-Python sebenzisa from verispoof import VeriSpoof, bese model = VeriSpoof.from_pretrained("AustinXiao/VeriSpoof-Plus") ukuze uqedele ukulayisha.
Ukutholwa kobuqiniso: shayela model.detect("sample.wav") ukuze ubuyise ilebula nesikolo. Isikolo sokutholwa nendawo yamathokheni ngamanye kudinga i-forward pass eyodwa kuphela, ngaphandle kokukhiqiza umbhalo.
Ukucabanga okuqinisekisekayo: shayela model.explain("sample.wav") ukuze ubuyise umbhalo wokucabanga, isiphetho, izingxenye ezibekwe endaweni kanye namathuba okwenziwa ngokwenziwa ngethokheni ngayinye. Ukucabanga okuphelele kungakhiqizwa ngobuningi nge-backend ye-vLLM; ku-A100, kuthatha cishe imizuzwana engu-0.05 ngento ngayinye.
Ukuqinisekisa umsuka ofanayo: shayela model.same_source("a.wav", "b.wav", level="generator") ukuze unqume ukuthi ukuqoshwa okubili kunomsuka ofanayo yini. I-level ingasethwa futhi ibe yi-acoustic_model noma i-vocoder.
Ukuhlonza izimo nokulandelela umthombo: shayela model.condition("sample.wav", attr="codec") ukuze uhlonze umsindo, i-reverberation noma i-codec; shayela model.generator("sample.wav", ["XTTS-v2", "Bark", "VITS"]) ukuze uhlonze uhlelo lokukhiqiza ohlwini lwamakhandidethi olwenziwe ngokwezifiso.

Ukusebenza okuphezulu phakathi kwezinhlelo ezinomthombo ovulekile: i-EER ehlanganisiwe ku-Speech-DF-Arena ingu-2.14%, ikleliswe endaweni yesithathu phakathi kwezinhlelo ezingu-28 futhi ingeyokuqala ngezinsimbi ezinomthombo ovulekile. Iphinde ibe ukuphela kohlelo oluhamba phambili olunohlelo lokuphindaphinda olungabizi kakhulu.
Isakhiwo esisodwa esikhundleni samamodeli amahlanu: izisindo ezifanayo zenza ukutholwa, ukulandelela umthombo, ukuqinisekisa umsuka ofanayo, ukuhlonza isiteshi nokucabanga; ayikho i-classification head eqondene nomsebenzi.
Isikolo siyimpendulo: zonke izikolo zokuhlola zivela ngokuqondile ekungenzekeni kwe-log kwemisho yezimpendulo zemodeli. Akukho ukuhlukana lapho i-classification head inamandla kodwa impendulo ingahambisani nombuzo; isiphetho esibonwa ngumsebenzisi yiso kanye esithole isikolo.
Ukucabanga kungaqinisekiswa, akusikho ukuqamba kokukhiqiza: ukucabanga kubhalwa ngezinkambu ezimisiwe, futhi umugqa ngamunye ungahlolwa ngokuqhathaniswa nezilinganiso ze-waveform noma amalebula edathasethi. Ukuqeqeshwa okuphikisanayo kuqinisekisa ukuthi isiphetho siguquka kuphela ngokuthi “kuphi okwenziwe ngokwenziwa”; izinga lokwethembeka lifinyelela ku-100%.
Ikhono lokwenza kahle ezimweni ezintsha liqinisekiswe ngokuqinile: ekwahluleleni umsuka ofanayo kumajeneretha angabonwanga, i-EER ingu-6.2%, ephansi ngo-77% kuneye-probe yama-pair ezinikele. Amajeneretha angabonwanga akhishwe ngokuphelele kudatha yokuqeqesha, ngakho umthetho wokuhlola uhlanzekile.

Iwebhusayithi esemthethweni yephrojekthi: https://swagshaw.github.io/verispoof/
Ikhosombe le-GitHub: https://github.com/swagshaw/verispoof
Ikhosombe lamamodeli le-HuggingFace: https://huggingface.co/collections/AustinXiao/verispoof
Iphepha lobuchwepheshe le-arXiv: https://arxiv.org/pdf/2611.04634

Ukumodareyitha kwemidiya nezinkundla zokuxhumana: ihlola ukuthi umsindo olayishwa abasebenzisi kanye “nezithombe-skrini zezwi” kwenziwe yi-AI yini, isize ukuvimba okuqukethwe okungamanga okwenziwe ngamazwi akopishiwe.
Ukuvimbela ukukhwabanisa kwezezimali nezingcingo: iqinisekisa ukuthi izingcingo zesevisi yamakhasimende nemiyalelo yokudlulisa imali iqoshwe ngumuntu wangempela yini, ivikele ezimweni zokukhwabanisa lapho izwi elikopishiwe lisetshenziswa ukuzenza isihlobo noma umphathi omkhulu.
Ubufakazi bezomthetho nokuphoqelelwa komthetho: inikeza imibiko yokucabanga engaqinisekiswa umugqa ngomugqa ekucutshungulweni kobufakazi bezwi, ichaze indawo yezingxenye ezikhohlisiwe nezinkomba zohlelo lokukhiqiza, futhi isize ekuhloleni komuntu.
Ukuqinisekisa ubuwena ezikhungweni zezingcingo: inquma ukuthi ukuqoshwa okugcinwe kanye nocingo lwesikhathi sangempela kunomsuka ofanayo yini nokuthi kuyizwi lomuntu wangempela yini, yehlise ingozi yokudlula ukuqinisekiswa kwezwi ngomsindo owenziwe ngokwenziwa.
Ucwaningo nokuhlolwa kwamamodeli: iyisisekelo esinamandla kakhulu sokutholwa phakathi kwezisindo ezinomthombo ovulekile kanye nemodeli yophenyo ehlanganisiwe, evumela abacwaningi ukuthi basebenzise i-VeriSpoof-Bench ukuhlola amamodeli abo nokwenza izivivinyo zokuqhathanisa ezilawulwayo.