Overview
Moshi ndeye yakavhurika-sosi, chaiyo-nguva izwi AI kubva kuKyutai inotaura uye inoteerera panguva imwe chete - yakazara-duplex - pane kutora majana akasimba. Izvo zvinobvisa kusadzikama uye kuomarara-kutora kwechinyakare vabatsiri vezwi.
Moshi Yakazara-Duplex Kutaura inogara muodhiyo-AI workflows inoshandura kutaura, mimhanzi, uye ruzha rwekutaurirana, kuwanikwa, uye kugadzirwa kwenhau.
Deep Dive
Moshi, yakaburitswa neFrench lab Kyutai muna 2024, imhando yekutaura-ku-kutaura-nheyo yakavakirwa kune yakasikwa, yakaderera-latency hurukuro. Kusiyana nemapaipi vabatsiri vanobatanidza kutaura-kune-mavara, ipapo modhi yemutauro, ipapo mavara-kune-kutaura, Moshi inobata odhiyo zvakananga uye ichienderera. Pfungwa yayo yakakosha ndeye duplex yakazara: inoenzanisira maodhiyo hova maviri panguva imwe chete - yemushandisi neyayo - saka inogona kuteerera ichitaura, inobata zvinokanganisa, backchannel ine 'mhm,' uye inopindirana zvakasikwa sezvinoita vanhu. Iyo inosvika latency yakatenderedza 160-200 milliseconds, kure pasi peyakajairwa mubatsiri lag. Pasi pehodhi inobatanidza 7B-parameter mavara-uye-odhiyo mutauro modhi (Helium) naMimi, neural audio codec inomanikidza kutaura kuita ma tokeni akasarudzika anogona kuburitswa nemuenzaniso. Kyutai akaburitsa huremu nekodhi pachena.
Technical Insight
Hunyengeri hwaMoshi ndeye Mimi codec, iyo inoshandura inoenderera odhiyo kuita yakaderera-bitrate rukova rwe discrete tokeni pa12.5 Hz, kusanganisira distilled semantic tokeni. Mutevedzeri wemutauro unofanotaura zviratidzo zvayo zvekutaura uye zvemushandisi mune hova dzinoenderana nenguva, saka chizvarwa hachimbofaniri kumira kuti 'chiteerere.' Nzira ye'Inner Monologue' inofanotaura zvinyorwa zvisati zvaitika, zvichivandudza mhando yemitauro uye kubatana kwezvinotaurwa naMoshi.
Kudzidzira Moshi Yakazara-Duplex Kutaura
Moshi ndeye yakavhurika-sosi, chaiyo-nguva izwi AI kubva kuKyutai inotaura uye inoteerera panguva imwe chete - yakazara-duplex - pane kutora majana akasimba. Izvo zvinobvisa kusadzikama uye kuomarara-kutora kwechinyakare vabatsiri vezwi. Moshi Yakazara-Duplex Kutaura inogara muodhiyo-AI workflows inoshandura kutaura, mimhanzi, uye ruzha rwekutaurirana, kuwanikwa, uye kugadzirwa kwenhau. Kuti uvake kunzwisisa kwakadzama, bata Moshi Yakazara-Duplex Kutaura semuenzaniso wekushandisa, kwete chinhu chimwe chete: tsanangura zvinodiwa, kujekesa fungidziro, uye patsanura izvo zvinogona kuitwa nehurongwa hwakavimbika kubva kune zvichiri kuda kutonga kwenyanzvi.
Mukuita, zvikwata zvakasimba zvinoshandisa Moshi Yakazara-Duplex Kutaura zvinobata mhando, 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
Mufambidzani wezwi risina maoko iwe unogona kukanganisa pakati pemutsara, nemhinduro dziri pasi pe200 milliseconds.
Vhura hwaro hwekutsvagisa yekudzidza-chaiyo-nguva, yakazara-duplex inotaurwa nhaurirano isina mabhokisi matema evaridzi.
Kuwanikwa vabatsiri vanokurukura zvinyoro-nyoro nevashandisi vanoda nekukurumidza, zvakasikwa kumashure-uye-mberi.
Prototyping inovhiringidza mutengi-sevhisi izwi bots iyo backchannel uye inopindura munhu anenge afona achiri kutaura.
Maitiro Ekuita
Moshi Yakazara-Duplex Kutaura mukuita
Mufambidzani wezwi risina maoko iwe unogona kukanganisa pakati pemutsara, nemhinduro dziri pasi pe200 milliseconds.
Mufambidzani wezwi risina maoko iwe waunogona kukanganisa pakati pemutsara, nemhinduro dziri pasi pe200 milliseconds Matimu anowanzo kuwana mhedzisiro iri nani kana achinge atsanangura maburi emhando kumberi, chengetedza nzira yekukwira kwevanhu yemakesi ekumucheto, uye kuteedzera zvese zvakawanikwa zvechigadzirwa uye mutengo wekukanganisa nekufamba kwenguva.
Moshi Yakazara-Duplex Kutaura mukuita
Vhura hwaro hwekutsvagisa yekudzidza-chaiyo-nguva, yakazara-duplex inotaurwa nhaurirano isina mabhokisi matema evaridzi.
Vhura hwaro hwekutsvagisa yekudzidza-chaiyo-nguva, yakazara-duplex inotaurwa nhaurirano isina evaridzi mabhokisi matema Matimu anowanzo kuwana mhedzisiro iri nani kana achinge atsanangura emhando yepamusoro kumberi, chengetedza nzira yekukwira kwevanhu yemakesi emupendero, uye kuteedzera zvese zvakawanikwa zvechigadzirwa nemitengo yekukanganisa nekufamba kwenguva.
Moshi Yakazara-Duplex Kutaura mukuita
Kuwanikwa vabatsiri vanokurukura zvinyoro-nyoro nevashandisi vanoda nekukurumidza, zvakasikwa kumashure-uye-mberi.
Vanowanika vabatsiri vanokurukura zvine mutsindo nevashandisi vanoda kukurumidza, zvisikwa kumashure-uye-kunze Matimu anowanzo kuwana mhedzisiro iri nani kana vachitsanangudza zvikumbaridzo zvemhando yepamusoro, chengetedza nzira yekukwira kwevanhu yemakesi emupendero, uye kuteedzera zvese zvakawanikwa zvechigadzirwa uye mutengo wekukanganisa nekufamba kwenguva.
Moshi Yakazara-Duplex Kutaura mukuita
Prototyping inovhiringidza mutengi-sevhisi izwi bots iyo backchannel uye inopindura munhu anenge afona achiri kutaura.
Prototyping inovhiringa mutengi-sevhisi izwi bots iyo backchannel uye inoita munhu anenge achifona achiri kutaura Matimu anowanzo kuwana mhedzisiro iri nani kana achinge atsanangura mhando yepamusoro kumberi, chengetedza nzira yekukwira kwevanhu yemakesi ekumucheto, uye kuteedzera zvese zvakawanikwa zvechigadzirwa nemitengo yekukanganisa 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.