Yeqa uye kokuyinhloko
Amabucwaningo omkhiqizo

I-OpenAI yethula amamodeli amabili okubhala okushiwo: ukubambezeleka okuncane nokuqonda kangcono umongo

Inkampani i-OpenAI imemezele namuhla (29 Julayi) ku-X ukuthi yethule amamodeli amabili okubhala okushiwo, i-GPT-Live-Transcribe ne-GPT-Transcribe, futhi ayatholakala ukuze asetshenziswe nge-API.

I-OpenAI yethula amamodeli amabili okubhala okushiwo: ukubambezeleka okuncane nokuqonda kangcono umongo

Inkampani i-OpenAI imemezele namuhla (29 Julayi) ku-X ukuthi yethule amamodeli amabili okubhala okushiwo, i-GPT-Live-Transcribe ne-GPT-Transcribe, futhi ayatholakala ukuze asetshenziswe nge-API.

I-OpenAI yethula amamodeli amabili okubhala okushiwo: ukubambezeleka okuncane nokuqonda kangcono umongo

Mayelana nezindleko zokusetshenziswa

Izindleko zokubhala okushiwo nge-GPT-Live-Transcribe ngu-0.017 USD ngomzuzu (Qaphela: ngezinga lokushintshisana lamanje cishe lilingana no-0.12 yuan wase-China)

Izindleko zokubhala okushiwo nge-GPT-Transcribe ngu-0.0045 USD ngomzuzu (ngezinga lokushintshisana lamanje cishe lilingana no-0.03 yuan wase-China)

I-OpenAI ithe ngok तुलना namamodeli ayo angaphambilini, la mamodeli amabili okuhumusha inkulumo abe umbhalo aqonda umongo kangcono futhi anikeza ukubhala okushiwo okunembe kakhulu kumsindo womhlaba wangempela onezindlela zokuphimisa nezilimi ezihlukahlukene, okuhlanganisa imishwana, izinombolo, amagama ochwepheshe kanye nenkulumo enomsindo omkhulu ongemuva.

I-GPT-Live-Transcribe yakhelwe ngokukhethekile ukubhala okushiwo ngesikhathi sangempela ngokubambezeleka okuncane; kanti i-GPT-Transcribe yenzelwe ukuthuthukisa ukubhala okushiwo okungavumelani ngesikhathi samafayela omsindo aqediwe kanye nemithwalo yomsebenzi yenqwaba.

I-OpenAI yethula amamodeli amabili okubhala okushiwo: ukubambezeleka okuncane nokuqonda kangcono umongo

Ekuhlolweni okuyisisekelo kwe-Context Aware ASR, ukunemba kwencazelo kwe-GPT-Transcribe kukhuphuke kusuka ku-41.6% ngaphandle komongo okhululekile kuya ku-45.2% lapho kukhona umongo.

I-OpenAI yethula amamodeli amabili okubhala okushiwo: ukubambezeleka okuncane nokuqonda kangcono umongo

Ezilimini ezingu-22 ze-Common Voice, izinga lamaphutha okubhala okushiwo kwe-GPT-Transcribe lingu-19.27%, kanti ele-Whisper (whisper-1) lingu-40.37%. Ezilimini eziyisishiyagalolunye ekuhlolweni okuyisisekelo kokuqoshwa komsindo womhlaba wangempela, ithole izinga lamaphutha okubhala okushiwo elingu-8.98%, kanti ele-Whisper (whisper-1) lingu-15.21%.