Jasper ak ASR
Jasper ak QuartzNet ñooy xeetu xàmmee kàddu yiñ defaree NVIDIA, QuartzNet mooy xeetu Jasper bi gëna ndaw te baax.
Résumé
They matter for showing how to get strong accuracy with far fewer parameters, ideal for deployment.
Plongeur bu xóot
Jasper (Beneen Xammekatu Kaddu rek), bi NVIDIA genne ci 2019, reso convolutionnel 1D bu xóot la, ba 54 diisaay, buy màndargaal màndarga mel-spectrogram ci araf yiy jëfandikoo CTC. Dugalal na lëkkaloo yu dëgër yu des suko defee gradient yi di naaw bu baax ci biir piil yu xóot yi. QuartzNet, bi ñu génne ci at moomu, denc na jumtukaayu bloku Jasper waaye mu wecci convolution yiñ miin ak convolution yuñ mëna tàqale ci chaine temps, ñu xaaj filtre bu nekk ci convolution temporel bu xóot ak jéego buy jaxase chaine pointwise. Factorisasioŋ bi dafa wàññi parametre yi ci Jasper ci lu tollu ci 333 milioŋ wàcci ba ci lu tollu ci 19 milioŋ ci noonu mu méngoo ak njubte ci Librispeech. Ñoom ñaar ñépp ñu ngi nekk ci jumtukaayu NeMo bu NVIDIA, ñu ngi leen di defar ngir tàggat GPU bu gaaw ak gis-gis ci jamono dëgg, loolu moo leen def ay bloku tabax yu siiw ngir defar ASR.
Gis-gis xarala
Xarañteg QuartzNet mingi aju ci convolution yuñ mëna tàqale ci chaine de temps, muy xalaat bi ci ginaaw MobileNet. Convolution 1D bu jaar yoon dafay jaxase jot ak chaine, muy njëg K yoon C-in yoon C-out poids. Tàqale ko ci benn convolution bu xóot ci diir bi boole ci benn convolution pointwise 1x1 ci kaw chaine yi dafay wàññi parametre yi ci K yoon C boole ci C-in yoon C-out. Ñu ngi ko dajale ci ay blok yu des, ñu tàggat ko ci CTC, loolu dafay joxe njubte bu jege Jasper ci benn wàll ci dayo model bi ak xayma.
njeextalu pexe
Dugg ak yegg
Dafay gëna yombal jëfandikoo gi jaaraleko ci transkripsioŋ, nettali ak interfaasu baat.
Njëgg ak budget
Ekipu mejaa yi mën nañu yónnee audio bu leer ci anam wu gëna gaaw te seen xaalis gëna néew.
Gaawaay ak yaatuwaay
Sistem yiy jàkkarloo ak kiliyaan bi mën nañu def waxtaan ci anam wu gëna yaatu.
Ëlëgu Jasper ak ASR
Liggéeyukaay yiñ mëna tàqale ci QuartzNet moo jur Citrinet bu NVIDIA ak xeetu Conformer yiñ gëna jëfandikoo, ñoom ñooy yokk seen xel ci seen bopp ngir jàpp jamonoy àdduna bi ci wetu convolution yu bërëb yi. Xaarandil dem bi ak dikk bi ci arsitektir hybrid ak transduser (RNN-T) decodeur ngir streaming. Njàngale mu am solo mi, convolutions parametre-efficace ngir edge ak deployment ci jamono dëgg, mingi des ci digg bi ASR di push ci telefon yi, oto yi ak aparey yiñ samp.
Doxal ci àdduna dëgg
Transkripsioŋ ci jamono dëgg ak jàppalekatu baat yuñ dugal ci GPU NVIDIA jaaraleko ci jumtukaayu NeMo
Edge ak ASR buñ samp fu emprent bu ndaw bu QuartzNet mëna méngoo ak aparey yu am memory bu néew
Defar bu baax barabi saytu QuartzNet yiñ tàggat bu njëkk ngir vocabulaire yu jëm ci domen bi lu melni terme medikaal wala yoon
Analytics call-center dafay bind audio yu bari ci anam wu gaaw te yomb
Risk yi ak balustrade yi
Jëfandikoo baat ci anam wu jaarul yoon ak niru ak nit dafay gëna yokk sudee nanguwul.
Jaar-jaar mën na wàññeeku ci aksan yi, dialect yi wala barab yu bari xumbaay.
Audio synthetik mën nañu ko jaawale ak wax ju dëggu sudee amul etiket bu leer.
Roadmap ngir samp gi
Wutal ndigal bu leer ngir jàpp baat bi, klone ko ak jëfandikoowaat ko.
Saytu kalite ci kàddukat yu bari ak anam yu bari ci ginaaw.
Mandargal kañ la nit wara xoolaat wala nangu ay génne.
Etiketu audio synthetik te nga denc dokimaa ci fimu bawoo ngir mëna lim.
Weyal di banneexu
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 Jasper and QuartzNet ASR 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
Gis bi ci topp
Wav2Letter ASR
Laaj yi ñuy faral di laaj
What is Jasper and QuartzNet ASR?
Jasper ak QuartzNet ñooy xeetu xàmmee kàddu yiñ defaree NVIDIA, QuartzNet mooy xeetu Jasper bi gëna ndaw te baax. Dañu am solo ngir wane ni ñuy amee njubte gu dëgër ak paramet yu néew, lu baax ci jëfandikoo gi.
Ban jéego bu am solo moo may QuartzNet mu jëfandikoo ay parametre yu gëna néew Jasper?
QuartzNet dafay wecci convolution yiñ miin ak convolution yuñ mëna tàqale ci chaine temps, muy wàññi bu baax limu parametre yi boole ci baña gaawa jeex.
Lu tollu ci ñaata paarametre la QuartzNet am buñu ko méngale ak Jasper ~333 milioŋ?
QuartzNet dafa wàññeeku ba tollu ci 19 milioŋ ciy parametre, lu tollu ci 17x wàññeeku, ci noonu mu méngoo ak dëggug Librispeech bu Jasper.
Ban liggéeyukaay moo defar Jasper ak QuartzNet?
Ñari xeetu model yooyu yépp NVIDIA moo leen defar, ñu ngi leen séddale ci jumtukaayu IA bu NeMo biy waxtaan.
Ban fonction perte la Jasper ak QuartzNet di jëfandikoo ngir tàggat te amul benn yoon bu leer?
Ñoom ñaar ñépp ñu ngi jëfandikoo CTC loss, muy méngale audio bu am guddaay bu mën soppiku ak seeni araf te soxlawul etiketu kaadar bu nekk.
Ban màndarga jëmmal mooy jàppale gradient yi ñu jaar ci reso bu xóot bi (ba 54-layer) bu Jasper?
Jasper dafay jëfandikoo lëkkaloo yu dëgër (sal) suko defee gradient yi di tasaaroo bu baax ci biir stack convolutionnel bu xóot bi.