SafeVoice — injini yobuhlakani benkulumo yobunikazi ekhombisa ukudakwa okungaba khona kusuka emasekhondini ambalwa nje enkulumo evamile. Akuhlaseli, akukho hardware ekhethekile, kudingeka nje ibinzana elikhulunyiwe.
Abasebenzi abadakiwe bahlala beyimbangela ehamba phambili futhi engahlolwanga kahle yezigameko zezimboni ezingavinjelwa. Utshwala bwehlisa isivinini sokusabela, ukwahlulela kwendawo, nokulawula imisipha emincane ngaphambi kokuba izimpawu zibonakale. Izilawuli zendabuko — ukuhlolwa kwe-breathalyzer ngezikhathi ezithile nokuhlolwa okubukwayo komqaphi — ziyashintshana, ziyahlasele, ziyizithiyo emasangweni ezishifti, futhi kulula ukuzigwema. Umphumela wukungaboni kahle kokusebenza: izimo ezinobungozi obukhulu zihlala zinzima kakhulu ukuzibamba ngokuqhubekayo.
Izwi liyisignali ekahle yokuqeda leli gebe. Abasebenzi basebenzisa ngokwemvelo ama-radio e-Push-to-Talk, bafinyelela ama-intercom, nezixhumi ezibonakalayo zezwi kuwo wonke umsebenzi wabo. Utshwala buguqula ngokuqhubekayo i-biomechanics yomsindo — kuphazamisa isivinini sokuphimisa, isikhathi se-phonation, i-prosody, nekhwalithi yezwi le-spectral — ama-biomarkers imodeli ekhethekile ewakhipha kumasekhondi omsindo.
I-SafeVoice isatshalaliswa ngezindlela ezimbili ezihambisanayo.
I-SafeVoice-Ref iqhathanisa izwi lesikhathi sangempela nesisekelo somsebenzi esibhalisiwe sokungadaki — elungele ukuhlolwa kokufinyelela ekuqaleni kweshifti.
I-SafeVoice-Solo isebenza ngaphandle kwesisekelo esibhalisiwe, ihlola ukuqhuma kwenkulumo okukodwa phakathi nokuxhumana kwe-radio okujwayelekile kuwo wonke umsebenzi.
Ukuhlukaniswa kokuhlolwa okugodliwe kwe-German Alcohol Language Corpus, ibenchmark yomphakathi ejwayelekile, kukalwe nge-unweighted average recall (UAR). Ukuqhathaniswa nobuciko bangaphambili kusebenzisa imethrikhi ye-per-window — isinqumo esisodwa nge-utterance ngayinye, akukho kuvota — ngoba ubuciko bangaphambili abuvoti.
Esikhundleni sokuhlukanisa umsindo uwedwa, i-SafeVoice-Ref iqhathanisa inkulumo nephrofayela yesisekelo yomuntu ngamunye. Ukuqeda ukwehluka kwezwi phakathi kwabakhulumi kunika ukukhula okusheshayo cishe ngamaphuzu angu-5 ku-Unweighted Average Recall (UAR) — ingqophamlando yethu yezakhiwo enomthelela omkhulu.
Kokubili i-Ref ne-Solo zisebenzisa izethulo ezifanayo ze-acoustic ezisezingeni eliphansi. I-SafeVoice-Ref ikhulisa ukunemba ngokubhaliswa okukodwa okusheshayo, kuyilapho i-SafeVoice-Solo ibeka phambili ukungqubuzani kokusebenza okunguziro, ihlola noma yimuphi umkhulumi ngokushesha ngaphandle kwedatha yomlando.
Umkhuba ojwayelekile uphakamisa ukuqandisa ama-backbone aqeqeshwe ngaphambilini kuma-corpora amancane. Ukuhlolwa kwethu okunamandla kufakazele okuphambene: ukunemba kokuhlolwa kukhula ngokujula kokulungisa, kugxuma cishe ngamaphuzu angu-7 ngaphansi kokulungiswa okugcwele kwe-backbone yonke uma kuqhathaniswa nokuqandisa kwezingxenye ezithile.
Ama-biomarker e-acoustic akhona ngezikhathi ezihlukene. Izendlalelo eziphansi ze-backbone zigcina izici ezinhle ze-articulatory ne-vocal tract, kuyilapho izendlalelo eziphezulu zibonisa ukugeleza kwe-prosodic nesigqi semisho. Ukuhlanganiswa okunezinga eliningi kulandelwa ukuhlanganiswa okunakile kukala ngokuguqukayo amafreyimu axilonga kakhulu.
Sidlule umsindo omhlophe wokwenziwa sakhetha ama-curricula abanzi omsindo wezimboni — imishini enzima, ukuduma kwezimoto, ingxoxo yesixuku, nokuduma kwesakhiwo. Lokhu kwavala igebe lokusebenza ezimeni ezinzima zensimu ngaphandle kokwehlisa ukunemba kwesisekelo esihlanzekile.
Ukuze kuvinjelwe ukuqhuma kwesikhashana noma ukuhlanza umphimbo ukuthi kuvuse ama-alamu angamanga, umugqa wokufaka ulinganisa amafasitela amaningi agqagqene nge-utterance ngayinye. Ukusebenzisa ukuvota ngokuvumelana kwengeza amaphuzu angu-1–2 e-UAR futhi kwenza ukutholwa kumelane ngendlela emangalisayo nokuqhuma komsindo wezwe langempela.
I-Consensus UAR (MV@5), esekelwe kumbandela / ngaphandle kwesisekelo. Ishicilelwe kufaka phakathi izimo lapho ukunemba kwehla khona.
SafeVoice ayikali izinga lotshwala egazini futhi ayikwazi ukusebenza njengesizathu sokususa isisebenzi emsebenzini. Iyithuluzi lokuhlola elisezingeni lokuqala elibonisa ukuthi ubani odinga ukunakwa, ukuze kuqinisekiswe ngezindlela ezimisiwe. Ukuqhathanisa okufanele akukhona ne-breathalyzer kodwa ngenye indlela esetshenziswayo namuhla: ukubhekwa ngumphathi, okusebenza cishe ezingeni lokulalela okungaqeqeshiwe.
Yonke imiphumela ivela kukhophasi eyodwa — i-German Alcohol Language Corpus, eqoshwe kuzikhulumi ezingu-162 ngeqiniso lomhlaba legazi nomoya wokuphefumula ngomuntu ngamunye. Iyinkomba eyisithenjwa yalomsebenzi, kodwa iyikhophasi eyodwa, futhi konke ukuqhathanisa okungenhla kwenziwa ekuhlukaniseni kwayo kokuhlola okugciniwe.
Sishicilela izimo lapho ukunemba kwehla khona kanye nezibalo ezihamba phambili, futhi sishicilela lokho esikwenqabile: imizwa esizayo kanye namakhanda okubuyisela emuva egazini lotshwala (umphumela omubi), ukuqandisa kwengxenye yomgogodla (cishe amaphuzu ayisikhombisa kubi kakhulu), ukushuna komkhawulo wesinqumo (ukulungiswa kwezimpawu okubiza amaphuzu angu-0.2), kanye nokubekwa kwesimo sesikhulumi semodeli engenankomba — inzuzo yangempela yamaphuzu angu-0.6, elahlwa ngoba yayizodinga ukubhaliswa komsebenzisi ngamunye futhi kubhubhise inzuzo ka-Solo yokungabizi nokungabizi.
Ukuhlolwa kokufaneleka komsebenzi ekuqaleni kweshifti: ukungena ngemvakusho okufushane okuqhathaniswa nesisekelo esibhalisiwe
Ukuqapha okuqhubekayo okungenzi lutho kumjikelezo womsakazo ojwayelekile, ngaphandle kokubhalisa
Iphawu elilinganiselwe elihanjiswa kumphathi ofanele ukuze kuthathwe isinqumo somuntu
Isebenza ku-CPU ngaphansi komzuzwana owodwa wokubambezeleka ngewindi ngalinye lokuhlaziya
Ihlehla kahle lapho umongo wedemografi ungatholakali
Ikhonjiswa ngeplathifomu ye-XENOM kanye nezinye izimpawu zokusebenza zesiza