Overview
Iyo RNN-Transducer (RNN-T) ndeyekutenderera-inoshamwaridzika yekuziva yekutaura dhizaini inogadzirisa hukuru hukuru hweCTC - kusakwanisa kwayo kutevedzera zvinoenderana pakati pezvinobuda tokeni. Inopa simba rakawanda re-pa-mudziyo 'live' yekuziva kutaura kwaunoshandisa mazuva ese.
RNN-Transducer Models inogara muodhiyo-AI workflows inoshandura kutaura, mimhanzi, uye ruzha rwekutaurirana, kuwanikwa, uye kugadzirwa kwenhau.
Deep Dive
Uyewo yakaunzwa naAlex Graves (2012), iyo RNN-Transducer inobatanidza zvikamu zvitatu. Iyo encoder (iyo transcription network) inogadzirisa maodhiyo mafuremu kuita acoustic maficha. Netiweki yekufembera inoita senge modhi yemutauro, inomisikidza pakutevedzana kweakamboburitswa mavara tokeni. Iyo diki network yakabatana inozobatanidza maonero eencoder ekuti 'patiri muodhiyo' nefungidziro yetiweki ye 'zvatataura kusvika parizvino' kuti tipe chiratidzo chinotevera pamusoro pemazwi anosanganisira isina chinhu. Kusiyana neCTC, network yekufungidzira inobvisa fungidziro-yekuzvimiririra, saka RNN-T inodzidza zviperengo zvechokwadi uye mapatani emazwi mukati. Decoding inofamba 2D lattice yeaudio-nguva inopesana neinobuda-tokens, ichiburitsa mablanks kuti ifambire mberi kuburikidza neodhiyo uye chaiyo tokens kufambira mberi kuburikidza nemavara - sechisikigo inotsigira kuburitsa kuburitsa.
Technical Insight
Kurasikirwa kweRNN-T, senge CTC's, inokwana pamusoro penzira dzese dzakasimba dzekugadzirisa kuburikidza nekudzoka-kumashure kudzokororwa, asi pamusoro pegidhi-mativi maviri (matanho enguva nenzvimbo dzekubuda) pane kutevedzana kumwe chete. Kuburitsa isina-isina kuvharika inogara pane imwechete odhiyo furemu uye inosimudzira iyo label index; kubudisa nguva yekufambira mberi isina chinhu. Iyi monotonic, yekuruboshwe-ku-kurudyi chimiro ndicho chikonzero nei RNN-T ichiyerera zvakachena ine bounded latency, kusiyana nekutarisa kuzere uko kunogona kutarisa pakutaura kwese.
Mastering RNN-Transducer Models
Iyo RNN-Transducer (RNN-T) ndeyekutenderera-inoshamwaridzika yekuziva yekutaura dhizaini inogadzirisa hukuru hukuru hweCTC - kusakwanisa kwayo kutevedzera zvinoenderana pakati pezvinobuda tokeni. Inopa simba rakawanda re-pa-mudziyo 'live' yekuziva kutaura kwaunoshandisa mazuva ese. RNN-Transducer Models inogara muodhiyo-AI workflows inoshandura kutaura, mimhanzi, uye ruzha rwekutaurirana, kuwanikwa, uye kugadzirwa kwenhau. Kuti uvake kunzwisisa kwakadzama, bata RNN-Transducer Models semuenzaniso wekushandisa, kwete chinhu chimwe chete: tsanangura zvinodikanwa, kujekesa fungidziro, uye patsanura zvinogona kuitwa nehurongwa hwakavimbika kubva kune izvo zvichiri kuda kutonga kwenyanzvi.
Mukuita, zvikwata zvakasimba zvinoshandisa RNN-Transducer Models zvinobata kunaka, latency, uye mvumo sezvikamu zvakakosha zvakaenzana zvehurongwa hwekuendesa. Ivo vanonyora zvakajeka maitiro ebudiriro, bvunzo vachipokana ne data rechokwadi uye mafambiro ebasa, uye iterate zvichibva pane zvakacherechedzwa maitiro ekutadza kwete kuhwina-nguva imwe chete yebhenji. Apa ndipo apo kunzwisisa kwe theoretical kunoshanduka kuve kugona kwakasimba pane chigadzirwa, mutemo, uye mashandiro.
Inonatsiridza kusvikika kuburikidza nekunyora, kurondedzera, uye mazwi ekubatanidza. Panguva imwecheteyo, kusashandiswa kweIzwi zvisizvo uye njodzi dzekuedzesera dzinowedzera kana chibvumirano chisipo. Nzira yakatsiga ndeyekubatanidza kukurumidza kuyedza nekutonga: mhanyisa vatyairi vendege, tora humbowo, buritsa matanda esarudzo, uye urambe uchivandudza chengetedzo semaitiro emuenzaniso, zvinotarisirwa nemushandisi, uye zvinodikanwa zvekutonga.
Strategic Impact
Inonatsiridza kusvikika kuburikidza nekunyora, kurondedzera, uye mazwi ekubatanidza.
Inonatsiridza kusvikika kuburikidza nekunyora, kurondedzera, uye mazwi ekubatanidza. Mukutumirwa kwemhando yepamusoro, izvi zvinoshandurirwa kuita mitemo inoyerwa yekushanda, miganhu yevaridzi, uye tsika dzekudzokorora dzinodzokororwa kuitira kuti zvikwata zvikwire kuvimba pane kukwidza kusajeka.
Zvikwata zveMedia zvinogona kutumira odhiyo yakakwenenzverwa nekukurumidza nemabhajeti madiki.
Zvikwata zveMedia zvinogona kutumira odhiyo yakakwenenzverwa nekukurumidza nemabhajeti madiki. Mukutumirwa kwemhando yepamusoro, izvi zvinoshandurirwa kuita mitemo inoyerwa yekushanda, miganhu yevaridzi, uye tsika dzekudzokorora dzinodzokororwa kuitira kuti zvikwata zvikwire kuvimba pane kukwidza kusajeka.
Masisitimu anotarisana nevatengi anogona kugadzirisa kutaurirana kwekutaura pamwero mukuru.
Masisitimu anotarisana nevatengi anogona kugadzirisa kutaurirana kwekutaura pamwero mukuru. Mukutumirwa kwemhando yepamusoro, izvi zvinoshandurirwa kuita mitemo inoyerwa yekushanda, miganhu yevaridzi, uye tsika dzekudzokorora dzinodzokororwa kuitira kuti zvikwata zvikwire kuvimba pane kukwidza kusajeka.
Real-World Implementation
Google inocherechedzwa yekutaura pamudziyo yeGboard diction uye Pixel Recorder, inoshanda pasina Indaneti.
Tsanangudzo yepamoyo inofambisa mazwi paunenge uchitaura pane kumirira kuti upedze chirevo
Vabatsiri vezwi vanonyora mirairo ine low latency iwe uchiri kutaura
Musangano wenguva-chaiyo uye kunyorwa kufona uko mibairo isina kukwana inofanira kuoneka nguva dzose
Maitiro Ekuita
RNN-Transducer Models mukuita
Google yekutaura pamudziyo weGboard diction uye Pixel Recorder, inoshanda pasina Indaneti.
Google's-on-device speech recognition yeGboard dictation nePixel Recorder, inoshanda zvizere pasina Indaneti Matimu anowanzowana mibairo iri nani paanotsanangura mabhindauko emhando kumberi, chengetedza nzira yekukwira kwevanhu yemakesi ekupedzisira, uye kuronda zvose zvinobudirira kubudirira uye mutengo wekukanganisa nekufamba kwenguva.
RNN-Transducer Models mukuita
Tsanangudzo yepamoyo inofambisa mazwi paunenge uchitaura pane kumirira kuti upedze chirevo.
Live Captioning inoyerera mazwi paunenge uchitaura pane kumirira iwe kuti upedze chirevo Matimu anowanzo kuwana mhedzisiro iri nani kana achinge atsanangura emhando yepamusoro kumberi, chengetedza nzira yekukwira kwevanhu yemakesi emupendero, uye tarisa zvese zvakawanikwa zvechigadzirwa uye mutengo wekukanganisa nekufamba kwenguva.
RNN-Transducer Models mukuita
Vabatsiri vezwi vanonyora mirairo ine low latency iwe uchiri kutaura.
Vabatsiri vezwi vanonyora mirairo ine yakaderera latency iwe uchiri kutaura Matimu anowanzo kuwana mhedzisiro iri nani kana achinge atsanangura emhando yepamusoro kumberi, chengetedza nzira yekukwira kwevanhu yemakesi ekumucheto, uye tarisa zvese zvakawanikwa zvechigadzirwa uye mutengo wekukanganisa nekufamba kwenguva.
RNN-Transducer Models mukuita
Musangano wenguva-chaiyo uye kunyorwa kufona uko mibairo isina kukwana inofanira kuoneka nguva dzose.
Musangano wenguva-chaiyo uye kudhindwa kwekufona uko mhedzisiro inofanirwa kuenderera mberi Matimu anowanzo kuwana mhedzisiro iri nani kana achinge atsanangura emhando yepamusoro kumberi, chengetedza nzira yekukwira kwevanhu yemakesi ekumucheto, uye kuteedzera zvese zvakawanikwa zvechigadzirwa uye mutengo wekukanganisa nekufamba kwenguva.
Njodzi & Guardrails
Kusashandisa izwi zvisizvo uye njodzi dzekuedzesera dzinowedzera kana chibvumirano chisipo.
Kururama kunogona kudonha mumitauro, mataurirwo, kana nharaunda dzine ruzha.
Synthetic audio inogona kukanganisa kutaura kwechokwadi isina mavara akajeka.
Implementation Roadmap
Wana mvumo yakajeka yekutora inzwi, kugadzira, uye kushandisa zvakare.
Wana mvumo yakajeka yekutora inzwi, kugadzira, uye kushandisa zvakare. Bata nhanho yega yega segedhi rehumbowo: kana maitiro asina kusangana, imbomira kuburitsa, vhara gaka, uye wobva wawedzera kushandiswa.
Yedza mhando pavatauri vakasiyana uye mamiriro ekumashure.
Yedza mhando pavatauri vakasiyana uye mamiriro ekumashure. Bata nhanho yega yega segedhi rehumbowo: kana maitiro asina kusangana, imbomira kuburitsa, vhara gaka, uye wobva wawedzera kushandiswa.
Tsanangura apo munhu anofanira kuongorora kana kubvumidza zvabuda.
Tsanangura apo munhu anofanira kuongorora kana kubvumidza zvabuda. Bata nhanho yega yega segedhi rehumbowo: kana maitiro asina kusangana, imbomira kuburitsa, vhara gaka, uye wobva wawedzera kushandiswa.
Label synthetic odhiyo uye chengetedza marekodhi ekuzvidavirira.
Label synthetic odhiyo uye chengetedza marekodhi ekuzvidavirira. Bata nhanho yega yega segedhi rehumbowo: kana maitiro asina kusangana, imbomira kuburitsa, vhara gaka, uye wobva wawedzera kushandiswa.