SafeVoice — injini ya luntha la mawu yomwe imazindikira zizindikiro za mowa kuchokera pa masekondi ochepa a mawu wamba. Zosasokoneza, zopanda zida zapadera, zomwe zimangofunika mawu olankhulidwa.
Ogwira ntchito oledzera akhalabe chomwe chimayambitsa zochitika zambiri zomwe zitha kupewedwa m'mafakitale. Mowa umachepetsa kuthamanga kwa kachitidwe, kuweruza kwa malo, ndi kayendetsedwe kake kabwino kale kwambiri zizindikiro zisanawonekere. Njira zachikhalidwe — zoyezera mowa nthawi ndi nthawi ndi kuyang'ana kwa oyang'anira — ndi zapakatikati, zosokoneza, zotsekereza pa zipata za shifti, ndipo ndizosavuta kuzemba. Zotsatira zake ndi malo osawoneka bwino antchito: zinthu zowopsa kwambiri zimakhala zovuta kwambiri kuzizindikira mosalekeza.
Mawu ndi chizindikiro chabwino kwambiri chochotsa mpata uwu. Ogwira ntchito amagwiritsa ntchito mwachibadwa mawailesi a Push-to-Talk, ma intercom, ndi mawonekedwe a mawu pa nthawi ya shifti yawo. Mowa umasintha mosasintha kayendetsedwe ka mawu — kumachepetsa kuthamanga kwa katchulidwe, nthawi ya phonation, prosody, ndi mtundu wa mawu — zizindikiro zomwe mtundu wapadera umazindikira kuchokera pa masekondi a audio.
SafeVoice imagwira ntchito m'njira ziwiri zothandizana.
SafeVoice-Ref imafanizira mawu a nthawi yeniyeni ndi mzere wolembedwa wa wogwira ntchito wosaledzera — yabwino pakuyesa mwayi pa nthawi yoyamba ya shifti.
SafeVoice-Solo imagwira ntchito popanda mzere wolembedwa, ikusanthula mawu amodzi pa nthawi ya mawailesi wamba pa nthawi ya shifti.
Gawo loyesedwa losatulutsidwa la German Alcohol Language Corpus, benchmark yaboma yodziwika, yoyesedwa ndi unweighted average recall (UAR). Kufanizira ndi zolembedwa zakale kumagwiritsa ntchito metric ya pa windo — chiganizo chimodzi pa mawu, popanda kuvota — chifukwa zolembedwa zakale sizivota.
M'malo moika audio payokha, SafeVoice-Ref imafanizira mawu ndi mbiri ya munthu. Kuchotsa kusiyana kwa mawu pakati pa okamba kumapereka phindu lapafupifupi 5-point mu Unweighted Average Recall (UAR) — chipambano chathu chachikulu kwambiri pa kamangidwe.
Onse Ref ndi Solo amagwiritsa ntchito zomwezo zapansi za mawu. SafeVoice-Ref imakulitsa kulondola kudzera mu kulembetsa kumodzi kofulumira, pamene SafeVoice-Solo imayika patsogolo kusagwira ntchito kopanda zopinga, kuyesa wokamba aliyense nthawi yomweyo popanda mbiri yakale.
Njira yokhazikika imati aziimitsa ma backbone ophunzitsidwa kale pa ma corpus ang'onoang'ono. Kuyesa kwathu kunatsimikizira zosiyana: kulondola kwa kuyesa kumakula ndi kuya kwa kusintha, kukwera pafupifupi 7 points pa kusintha kwathunthu kwa end-to-end backbone kuyerekeza ndi kuzimitsa zigawo zina.
Zizindikiro za mawu zimapezeka pa nthawi zosiyanasiyana. Zigawo zapansi za backbone zimasunga kayendetsedwe kabwino ka mawu ndi katchulidwe, pamene zigawo zapamwamba zimayang'anira kayendedwe ka mawu ndi kayimba. Kuphatikiza kwamiyezo yambiri ndikutsatidwa ndi attentive pooling kumayika zolemba zofunika kwambiri.
Tidapewa phokoso loyera lopangidwa ndi kupanga m'malo mwake maphunziro ambiri a mafakitale — makina olemera, phokoso la magalimoto, phokoso la anthu, ndi kubwereza kwa nyumba. Izi zidatseka kusiyana kwa magwiridwe antchito pamikhalidwe yovuta popanda kuwononga kulondola kwa maziko oyera.
Kuti tipewe kukwera kwadzidzidzi kapena kutsokomola kuti kusayambitse ma alarm abodza, njira yowerengera imayesa mawindo angapo osiyanasiyana pa mawu aliwonse. Kugwiritsa ntchito kuvota kogwirizana kumawonjezera 1–2 points za UAR ndipo kumapangitsa kuzindikira kukhala kolimba kwambiri ku phokoso lenileni.
Consensus UAR (MV@5), yotengera mbiri / yopanda mbiri. Yosindikizidwa kuphatikiza mikhalidwe yomwe kulondola kumatsika.
SafeVoice siyeyesa kuchuluka kwa mowa m'magazi ndipo sichingagwiritsidwe ntchito ngati chifukwa chochotsera wogwira ntchito. Ndi chida choyamba chowunika chomwe chimasonyeza amene akuyenera kuyang'aniridwa, kuti atsimikizidwe ndi njira zokhazikitsidwa kale. Kufananitsa koyenera si ndi breathalyzer koma ndi njira ina yomwe ikugwiritsidwa ntchito lero: kuyang'anira kwa woyang'anira, komwe kumagwira ntchito pafupifupi monga kumvetsera kosaphunzitsidwa.
Zotsatira zonse zimachokera ku gulu limodzi — German Alcohol Language Corpus, lojambulidwa kuchokera kwa oyankhula 162 ndi zowona za mowa m'magazi ndi mpweya wa munthu aliyense. Ndi chizindikiro chofotokozera cha ntchito iyi, koma ndi gulu limodzi lokha, ndipo kufananitsa kulikonse pamwambapa kwapangidwa pa gawo lake loyesedwa losungidwa.
Timasindikiza mikhalidwe yomwe kulondola kumachepa pamodzi ndi ziwerengero zazikulu, ndipo timasindikiza zomwe tinakana: mitu yothandizira malingaliro ndi kubweretsa mowa m'magazi (zotsatira zoyipa), kusungitsa gawo la msana (pafupifupi mfundo zisanu ndi ziwiri zoyipa), kusintha kwa chisankho (kukonza chizindikiro kofunika mfundo 0.2), ndi kukonza mawu kwa wokamba nkhani kwa chitsanzo chopanda zofotokozera — phindu lenileni la mfundo 0.6, linatayidwa chifukwa likadafunikira kulembetsa kwa wogwiritsa ntchito aliyense ndikuwononga mwayi wa Solo wopanda zovuta.
Kuyesa koyenera kwa ntchito kumayambiriro kwa shifiti: kulowa mwachidule polankhula motsutsana ndi chizindikiro cholembetsedwa
Kuyang'anira kosalekeza kopanda chochita pa mawayilesi anthawi zonse, popanda kulembetsa
Chidziwitso choyezera chotumizidwa kwa woyang'anira woyenera kuti apange chisankho chaumunthu
Imagwira ntchito pa CPU ndi kuchedwa kwa mphindi imodzi pa zenera lililonse la kusanthula
Imachepa mwaulemu pamene nkhani za anthu sizikupezeka
Imawonetsedwa kudzera pa nsanja ya XENOM pamodzi ndi zizindikiro zina za ntchito za malowa