He aha ʻo Whisper AI? Ka hana o ka model ʻike leo a OpenAI

Na ka hui hoʻoponopono o whispertranscribe.ai · Hoʻohou hope ʻia: Nowemapa 2026 · 13 minuke heluhelu

He aha ʻo Whisper AI? Ma ka hapa nui o nā huli, ʻo OpenAI Whisper ia, he model ʻike leo open-source i paʻi ʻia e OpenAI i Kepakemapa 2022. Hoʻololi ia i ka ʻōlelo i hoʻopaʻa ʻia i kikokikona, ʻike i ka ʻōlelo, a hiki ke unuhi i nā ʻōlelo ʻē aʻe i ka Pelekānia. Ua aʻo ʻo OpenAI iā ia ma 680,000 hola leo mai ka pūnaewele a hoʻokuʻu i ka code a me nā weights ma lalo o ka MIT License.

He hōʻike pōkole: hana mākou iā whispertranscribe.ai. He mea AI transcription ma ka pūnaewele ʻo whispertranscribe.ai e hoʻololi ana i ka leo a me ke wikiō i kikokikona ma 130+ ʻōlelo. Hoʻokaʻawale kēia alakaʻi i ka model a me nā huahana i kūkulu ʻia ma luna, a mai ka pepa, repository a i ʻole kāleka model a OpenAI nā helu a pau e pili ana i ka model.

Pane pōkole: he aha ʻo Whisper AI?

He model speech-to-text ʻo Whisper, ʻaʻole he polokalamu a hui paha. Hiki iā ʻoe ke hoʻoiho mai ka GitHub repository a OpenAI a holo ma kāu kamepiula, a i ʻole hoʻohana i kahi lawelawe e holo ana no ʻoe.

  • Mea hana: OpenAI. Ua kau ʻia ka pepa noiʻi, Robust Speech Recognition via Large-Scale Weak Supervision, ma arXiv i Kekemapa 6, 2022.
  • Ka hana: hoʻololi i ka ʻōlelo, ʻike i ka ʻōlelo, unuhi i ka ʻōlelo i Pelekānia, a hoʻohui i nā manawa ʻāpana.
  • Nā mea i hoʻokomo ʻole ʻia: nā inoa mea ʻōlelo a me ka hoʻololi manawa maoli. Wahi a ke kāleka model a OpenAI, ʻaʻole hiki ke hoʻohana koke ʻia no ka transcription manawa maoli me ka hana hou ʻole.
  • Nā nui: ʻeono, mai tiny (39 miliona parameters) a i large (1,550 miliona).
  • Uku: manuahi ka hoʻoiho a holo ʻana. Na ʻoe nā lako kamepiula.

Hiki i “Whisper AI” ke manaʻo i ʻekolu mea ʻokoʻa

Hoʻohana ākea ʻia ka inoa, a maʻalahi ka huikau ʻana o kēia mau manaʻo ʻekolu.

Ka mea i kapa ʻia “Whisper AI” He aha ia Na wai i hana Pehea e hoʻohana ai
OpenAI Whisper He model ʻike leo open-source OpenAI Hoʻoiho i ka code a me nā weights a holo iā ʻoe iho
whisper-1 Ka model API i hoʻokipa ʻia e OpenAI no ka transcription OpenAI Hoʻouna i nā faila leo i ka API a uku no ka hoʻohana
Nā polokalamu me “Whisper” ma ka inoa Nā huahana transcription, e like me whisperai.com a me whispertranscribe.ai (kēia pūnaewele) Nā hui kaʻawale Hoʻouka a hoʻopaʻa leo ma kā lākou mau polokalamu

ʻAʻole ʻo nā huahana ma ka laina ʻekolu ka model ponoʻī a ʻaʻole hiki ke hoʻololi wale kekahi i kekahi. Hoʻonohonoho kēlā me kēia hui i nā hiʻohiʻona, nā palena, nā uku a me nā ʻōlelo pilikino. Ma kēia ʻatikala, ʻo “Whisper” ka model a OpenAI.

Pehea ka hana o Whisper?

He encoder-decoder Transformer ʻo Whisper. Hoʻololi ka encoder i 30 kekona leo i kahi ʻano helu, a kākau ka decoder i ka kikokikona pili i hoʻokahi token i ka manawa, alakaʻi ʻia e nā token kūikawā e haʻi ana i ka hana.

Hōʻike ka pepa Whisper i kēia kaʻina:

  1. Resample. Hoʻololi ʻia ka leo i 16,000 Hz.
  2. ʻOki i nā puka 30 kekona. Nānā mau ka model i 30 kekona i ka manawa. Hoʻopiha ʻia nā ʻāpana pōkole.
  3. Hana i spectrogram. Lilo kēlā me kēia puka i 80-channel log-Mel spectrogram, he kiʻi o ka loli ʻana o ka ikehu kani ma nā frequencies, i helu ʻia ma nā frame 25 millisecond i kēlā me kēia 10 millisecond. (Ua hoʻololi ʻo large-v3 i 128 channels.)
  4. Encode. Hana nā papa convolution ʻelua a me nā poloka Transformer i ka spectrogram.
  5. Decode. Wānana mua ka decoder i ka ʻōlelo a i ʻole ka token “no speech”. Ma hope ka hana (transcribe a translate), inā hoʻokomo i nā manawa, a laila ka kikokikona.

Wānana ʻia nā manawa he mau token pū, i hoʻopuni ʻia i ka 20 millisecond kokoke loa. Pēlā e mālama ai hoʻokahi model i ka ʻike ʻōlelo, transcription, unuhi a me ka manawa.

He aha ka mea i aʻo ai ʻo Whisper?

Ua aʻo nā model kumu mai 680,000 hola leo pūnaewele me nā transcripts pili. Kapa ʻo OpenAI he “weak supervision” no ka mea ʻaʻole i nānā lima ʻia nā transcripts e like me nā datasets hoʻāʻo.

Māhele o ka leo aʻo Nā hola Ka ʻike
65% paha 438,000 Leo Pelekānia me nā transcripts Pelekānia
18% paha 126,000 Leo ʻaʻole Pelekānia me nā transcripts Pelekānia (ʻikepili unuhi)
17% paha 117,000 Leo ʻaʻole Pelekānia me nā transcripts ma ka ʻōlelo like

Kumu: kāleka model Whisper a OpenAI. Hoʻopuni ka pepa i ka māhele unuhi i 125,000 hola.

Wehewehe ʻelua kikoʻī i ka ʻano o Whisper. Ua kānana ʻo OpenAI i nā transcripts e like me ka mea i hana ʻia e ka mīkini, e like me nā kikokikona huapalapala nui wale nō, no laila ua aʻo ka model e kākau me nā kaha kiko. Ua koho nā model mua i nā inoa mea ʻōlelo no ka mea nui nā transcripts pūnaewele i loaʻa ia mau inoa. Ua aʻo hou ʻo OpenAI e hōʻemi i ia ʻano, akā hōʻike ia i ka pili o ka model i kāna ʻike aʻo.

Ua aʻo ʻia ʻo large-v3 ma kahi ʻano ʻē aʻe, ma kahi o 5 miliona hola: 1 miliona me nā lepili pūnaewele nāwaliwali a me 4 miliona i lepili ʻia e large-v2, e like me kāna kāleka Hugging Face.

Hoʻohālikelike i nā nui model Whisper

ʻEono nui o Whisper. Holo wikiwiki nā model liʻiliʻi ma nā lako haʻahaʻa; emi nā hewa o nā model nui akā pono ka video memory (VRAM) hou aʻe. Mai ka README a OpenAI nā helu.

Nui Parameters Mana Pelekānia wale nō Mana ʻōlelo he nui VRAM i pono Wikiwiki me large
tiny 39 M tiny.en tiny ~1 GB ~10x
base 74 M base.en base ~1 GB ~7x
small 244 M small.en small ~2 GB ~4x
medium 769 M medium.en medium ~5 GB ~2x
large 1550 M ʻAʻohe large ~10 GB 1x
turbo 809 M ʻAʻohe turbo ~6 GB ~8x

Ua ana ʻo OpenAI i kēia mau wikiwiki ma ka leo Pelekānia me kahi A100 GPU kikowaena ʻikepili a ʻōlelo hiki ke ʻokoʻa nui ka wikiwiki maoli. Lohi aʻe ka laptop.

ʻO wai ka model Whisper e koho ai?

E hoʻomaka mai ka memory i loaʻa a me ka mea āu e pono ai:

  • Hoʻāʻo ma ke kamepiula kumu: tiny a base paha, ma kahi o 1 GB VRAM. E manaʻo i nā hewa hou aʻe.
  • Pelekānia wale nō: e hoʻohana i ka model .en. Wahi a OpenAI, maikaʻi aʻe kēia, ʻo tiny a me base hoʻi.
  • Maikaʻi me ka wikiwiki kūpono: turbo, ma kahi o 6 GB. ʻO ia ka paʻamau ma ka command line.
  • Unuhi i ka ʻōlelo Pelekānia: medium a large paha. ʻAʻole i aʻo ʻia ʻo turbo no ka unuhi a hoʻihoʻi ia i ka ʻōlelo kumu inā noi ʻoe e unuhi.
  • Leo paʻakikī loa, hopena maikaʻi loa: large, ma kahi o 10 GB VRAM.

Nā mana Whisper: he papa manawa pōkole

Ua komo ʻekolu hoʻokuʻu i loko o “large”, a he large-v3 i hoʻēmi ʻia ʻo turbo. He mea nui ka mana, no ka mea ʻokoʻa nā hopena ma ka checkpoint.

Mana Hoʻokuʻu ʻia Ka hoʻololi
Nā model kumu (tiny a large) Kepakemapa 2022 Hoʻokuʻu open-source mua me ka pepa
large-v2 Kekemapa 2022 Ka nui like, aʻo ʻia 2.5 manawa ka lōʻihi me ka regularization hou
large-v3 Nowemapa 2023 128 Mel channels, Cantonese token hou, 10% a 20% ka emi o nā hewa ma mua o v2 ma nā ʻōlelo he nui
turbo (large-v3-turbo) Kepakemapa 2024 Hoʻemi ʻia nā papa decoder mai 32 a i 4 no ka wikiwiki; ʻaʻohe unuhi

Nā kumu: ke kāleka model Whisper, ka footnote large-v2 o ka pepa, a me nā kāleka Hugging Face no large-v3 a me large-v3-turbo.

ʻEhia ʻōlelo e kākoʻo ai ʻo Whisper?

ʻIke ʻo Whisper i 99 ʻōlelo paha. Ua hoʻohana ka language detector o ka pepa i 99 language tokens, a papa inoa ka tokenizer o kēia manawa i 100 language codes ma hope o ka hoʻohui ʻana i Cantonese ma large-v3. ʻAʻole like ke kākoʻo me ka maikaʻi: wahi a ke kāleka model he hopena ikaika ma “~10 languages”.

Pili ka maikaʻi i ka ʻike aʻo. Ua ʻike ka pepa hiki ke wānana maikaʻi i ka hana o ka ʻōlelo ma ka nui hola i ʻike ʻia e ka model, a ma lalo o 1,000 hola ka hapa nui o nā ʻōlelo. ʻOi aku ka maʻemaʻe maʻamau o ka Pelekānia, Paniolo a Kelemānia paha ma mua o ka ʻōlelo me ka ʻikepili liʻiliʻi. Hoʻopili ka README i kahi pakuhi hewa ma kēlā me kēia ʻōlelo.

Mai hoʻohui i kēia me ka helu huahana. Kākoʻo ʻo whispertranscribe.ai i 130+ ʻōlelo. ʻO ia ke kākoʻo o kā mākou huahana (wahi a ka ʻaoʻao kumukūʻai “130+ languages & accents”), ʻaʻole he waiwai o ka model open-source Whisper.

Pehea ka pololei o Whisper?

ʻAʻohe helu hoʻokahi no Whisper, no ka mea pili ia i ka nui model, ka ʻōlelo a me ka leo. Ma ka Pelekānia maʻemaʻe i heluhelu ʻia, loaʻa iā large-v2 he 2.7% word error rate (WER). Ma ka awelika o nā datasets Pelekānia 14, ʻo ka hapa nui he paʻakikī aʻe, he 12.8%.

ʻO WER ka helu hewa o nā huaʻōlelo i pani ʻia, waiho ʻia a hoʻohui ʻia, i hoʻohālikelike ʻia me ka transcript kuhikuhi a ke kanaka. ʻOi ka haʻahaʻa. Mai ka papa 2 o ka pepa nā helu, e hoʻohālikelike ana iā Whisper large-v2 me kahi model wav2vec 2.0 i aʻo ʻia ma ka LibriSpeech audiobooks:

Dataset wav2vec 2.0 WER Whisper large-v2 WER
LibriSpeech clean (nā puke leo i heluhelu ʻia) 2.7% 2.7%
Common Voice (leo i hoʻopaʻa ʻia e ka lehulehu) 29.9% 9.0%
CHiME-6 (nā pāʻina ʻaina walaʻau) 65.8% 25.5%
Awelika o nā datasets 14 ma ka papa 29.3% 12.8%

ʻO ke ʻano ka mea nui. Like nā model ma ka hoʻāʻo puke leo, akā ʻoi aku ʻo Whisper ma nā mea ʻē aʻe, me ka emi 55.2% o nā hewa ma ka awelika. Hoʻopuni ka hoʻolaha OpenAI i “50% ka emi”. Ma nā hoʻāʻo walaʻau, ʻoi nā model i aʻo ʻia ma LibriSpeech ma ka leo kokoke maʻemaʻe, akā emi ma hope ke piʻi ka walaʻau hale inu ma lalo o 10 dB signal-to-noise ratio.

He mau ana noiʻi kēia, ʻaʻole he hoʻohiki no kāu mau faila, a kaʻawale nā ʻōlelo huahana. No ka laʻana, hiki iā whispertranscribe.ai ke hiki i 99% ma ka leo maʻemaʻe; hoʻemi ka walaʻau, nā leo e uhi ana a me nā leo kūloko ikaika.

He aha nā palena o Whisper?

ʻO ka nāwaliwali nui hiki iā ia ke kākau i ka ʻōlelo maikaʻi i ʻōlelo ʻole ʻia. Kapa ke kāleka model a OpenAI he hallucination a hoʻopili iā ia i ke aʻo weakly supervised. Papa inoa pū ia i ka hana hou, ka pololei like ʻole ma nā ʻōlelo a me nā leo kūloko, a me ka nele o nā inoa mea ʻōlelo i hoʻokomo ʻia.

  1. Kikokikona i haku ʻia. Ma ka noiʻi 2024 Careless Whisper, ua ʻike nā mea noiʻi he 1% paha o nā transcripts i nānā ʻia i loaʻa nā huaʻōlelo a ʻōlelo piha ʻaʻole i loko o ka leo. ʻOi pinepine ka hallucination me nā hoʻomaha lōʻihi.
  2. Nā loop hou. Hiki i ka model ke paʻa i ka hana hou ʻana i kekahi huaʻōlelo. Hoʻemi ka beam search a me temperature fallback i kekahi ʻāpana, akā ʻaʻole piha.
  3. Nā huaʻōlelo mua a hope nalowale. Papa inoa ka pepa i ka lele ʻana i nā huaʻōlelo mua a hope paha o ka ʻāpana he pilikia i ʻike ʻia.
  4. Pono nā puka no nā leo lōʻihi. ʻIke ʻo Whisper i 30 kekona i ka manawa. No nā faila lōʻihi, neʻe ka mea transcribe i ka puka me nā manawa i wānana ʻia, no laila hiki i ka hewa ke hoʻopilikia i ka puka aʻe.
  5. ʻAʻohe inoa mea ʻōlelo. ʻAʻole i loiloi paʻa ʻia ka diarization, ka ʻike ʻana i ka mea i ʻōlelo a me ka manawa, e like me ke kāleka. Pono kahi mea hana ʻē aʻe.
  6. ʻAʻole manawa maoli ma ka paʻamau. Hana ʻia no nā faila i hoʻopaʻa ʻia. Pono ka ʻenekinia hou no live captioning.
  7. ʻOkoʻa ma ka ʻōlelo a me ka leo kūloko. Hōʻike ke kāleka i nā helu hewa kiʻekiʻe aʻe no kekahi mau leo, dialect a me nā hui kanaka.

Aʻo pū ʻo OpenAI ʻaʻole e hoʻololi i nā leo kanaka me kā lākou ʻae ʻole a ʻaʻole e hoʻohana no nā hoʻoholo pilikia kiʻekiʻe.

Kahi e ʻokoʻa ai nā kumu kūhelu

ʻOkoʻa kekahi mau helu ma nā palapala ponoʻī o OpenAI, no ka hoʻopuni helu a me nā mana ʻokoʻa. E nānā ma mua o ka ʻōlelo ʻana:

Kumuhana ʻŌlelo a kekahi kumu kūhelu ʻŌlelo a kekahi Ka mea e hoʻohana ai
Ka emi hewa me nā model LibriSpeech Pepa: 55.2% ka emi ma ka awelika Hoʻolaha OpenAI: 50% ka emi E ʻōlelo i ka pepa a me ka hoʻohālikelike
Nā ʻōlelo aʻo ma waho o Pelekānia Pepa: 96 ʻōlelo ʻē aʻe Kāleka model: 98 ʻōlelo “99 ʻōlelo paha” no ka ʻike ʻana
Nā hola aʻo unuhi Kāleka model: 126,000 Pepa: 125,000 ʻO kekahi, me ka hōʻailona he kokoke
Nā parameters turbo README a me Hugging Face: 809 M Papa kāleka model: 798 M 809 M, ka mea i hoʻohana nui ʻia
Nui ʻikepili aʻo Pepa kumu: 680,000 hola Kāleka large-v3: 5 miliona hola E inoa i ka mana āu e ʻōlelo ai
Laikini GitHub: MIT no code a weights Metadata Hugging Face large-v3: Apache-2.0 E nānā i ka faila laikini o ke kope āu e hoʻoiho ai

Pehea e hoʻohana ai iā Whisper: holo iā ʻoe iho, API, a i ʻole polokalamu

ʻEkolu ala kūpono. Manuahi ka holo ʻana iā Whisper iā ʻoe iho a noho ka leo ma kāu kamepiula. Wehe ka API a OpenAI i ka hoʻonohonoho, akā palena ka nui faila. Hoʻohui kahi polokalamu hoʻokipa i ka hoʻoponopono, nā inoa mea ʻōlelo a me nā puka, akā he lawelawe ʻaoʻao ʻekolu me kāna mau ʻōlelo ponoʻī.

Holo Whisper open-source OpenAI transcription API Polokalamu hoʻokipa
Hoʻonohonoho Python, PyTorch a me FFmpeg Code a me API key Polokalamu kele pūnaewele
Uku Manuahi; kāu lako Uku no ka hoʻohana Pae manuahi a kau inoa, ʻokoʻa ma ka polokalamu
Nā palena faila Kau lako 25 MB no ka faila Hoʻonohonoho ʻia e ka polokalamu
Kahi e hele ai ka leo Noho ma kāu kamepiula Hoʻouna ʻia iā OpenAI Hoʻouna ʻia i nā server o ka polokalamu
Nā inoa mea ʻōlelo ʻAʻole i hoʻokomo ʻia ʻAʻole ma whisper-1 Pili i ka polokalamu
Hoʻoponopono a me nā puka TXT, SRT, VTT, TSV, JSON mai ka command line Hoʻi i kāu code; kūkulu ʻoe i ke koena Mea hoʻoponopono a me nā puka palapala a subtitle maʻamau

Holo iā Whisper ma kāu kamepiula

Hāʻawi ka README a OpenAI i ke kauoha hoʻokomo a me kahi holo kumu:

pip install -U openai-whisper
whisper interview.mp3 --model turbo

Pono pū ʻo FFmpeg, a ʻōlelo ka README e holo ka code ma Python 3.8 a i 3.11. E hoʻohui --language no ka ʻōlelo i ʻōlelo ʻia, a i ʻole --task translate me kahi multilingual model no ka Pelekānia. Hoʻoiho ʻia nā weights i ka hoʻohana mua.

E hoʻohana i ka API a OpenAI

Kau ka alakaʻi speech-to-text a OpenAI i ka palena 25 MB no ka faila, no laila e compress a māhele paha i nā mea nui aʻe. E like me ka nānā ʻana i Nowemapa 2026 o ka ʻatikala kumu, hōʻailona ke alakaʻi iā whisper-1 he deprecated a papa inoa i ka lā pani Pepeluali 26, 2027, a kuhikuhi i nā mea hoʻomohala i nā model transcription hou. He mau model leo OpenAI ia, ʻaʻole he hoʻokuʻu hou o ka ʻohana Whisper open-source.

E hoʻohana i kahi polokalamu hoʻokipa

Inā ʻaʻole makemake e hoʻokomo i kekahi mea, hana kahi polokalamu hoʻokipa i nā faila ma kāna mau server. ʻO whispertranscribe.ai kekahi koho, a he ala wikiwiki kāna mea hoʻololi leo manuahi e hoʻāʻo ai. Wahi a ka ʻaoʻao mua, hoʻohana ʻo whispertranscribe.ai iā Whisper a me AI hou. Eia nā hiʻohiʻona mai kā mākou mau ʻaoʻao huahana:

Aia nā mea e noʻonoʻo ai. Hele kāu faila i kahi cloud service ma kahi o ka noho ʻana ma kāu kamepiula. ʻAʻole mākou e hāʻawi i API, papa huaʻōlelo maʻamau a meeting bot paha. Hoʻomaka nā hoʻolālā uku ma $9.99/mahina ke uku makahiki ʻia (kumukūʻai), a hiki ke hoʻomaka manuahi me ka ʻole o ka moʻokāki. E ʻike i ka Whisper AI transcription tool a i ʻole nā hiʻohiʻona a pau.

He open source anei ʻo Whisper, a he aha ka mea i ʻae ʻia e ka laikini?

ʻAe. Wahi a ka repository a OpenAI, hoʻokuʻu ʻia ka code a me nā model weights o Whisper ma lalo o ka MIT License. He laikini ʻae ākea ka MIT. Ma ka laulā, ʻae ia i ka hoʻohana, kope, hoʻololi a kūʻai ʻana i ka polokalamu i kūkulu ʻia ma ka code, inā mālama ʻia ka hoʻolaha copyright a me ka laikini.

ʻElua mea e mālama ai. Hiki i nā kope ma nā pūnaewele ʻē aʻe ke loaʻa nā metadata laikini ʻokoʻa; papa inoa ka ʻaoʻao large-v3 ma Hugging Face iā Apache-2.0. E heluhelu i ka laikini o kāu kope. Uhi ka laikini i ka polokalamu, ʻaʻole i nā leo āu e hoʻokomo ai. Pili nō nā lula ʻae a pilikino i ka leo. He ʻike laulā kēia, ʻaʻole ʻōlelo aʻo kānāwai; ʻo ka kikokikona MIT ke kumu kūhelu.

No wai ʻo Whisper?

Kūpono ʻo Whisper no nā mea hoʻomohala a me nā mea ʻoluʻolu i ka terminal e makemake ana i ka transcription manuahi, pilikino a palena ʻole. Inā makemake ʻoe i kahi transcript piha me ka mea hoʻoponopono a me nā puka, ʻoi aku ka wikiwiki o ka polokalamu hoʻokipa.

ʻO ʻoe… Koho kūpono No ke aha
Mea hoʻomohala huahana Whisper open-source a API paha Hoʻomalu piha i ke kaʻina
Mea noiʻi me nā nīnauele koʻikoʻi Whisper open-source ma kāu kamepiula ʻAʻole haʻalele ka leo i kāu kamepiula
Mea hana e pono ai nā subtitles i kēia pule Polokalamu hoʻokipa me SRT a VTT ʻAʻohe hoʻonohonoho; mākaukau nā faila me ka manawa
Mea hoʻohana laptop me ka ʻole o GPU Model liʻiliʻi ma ke kamepiula, a i ʻole polokalamu hoʻokipa Pono nā model nui i 10 GB VRAM paha

Ma mua o ka hilinaʻi ʻana i ka transcript Whisper: papa hōʻoia

Ua kūkulu nā mea kākau o ka ʻatikala kumu i kēia papa mai nā pilikia ma ka pepa a me ke kāleka OpenAI:

  • E nānā i nā hoʻomaha a me nā hāmau no nā ʻōlelo i ʻōlelo ʻole ʻia.
  • E ʻimi i ka huaʻōlelo like i hana hou ʻia i nā manawa he nui.
  • E nānā i nā laina mua a hope o ka faila me ka leo.
  • E hōʻoia i nā inoa, nā helu a me nā huaʻōlelo ʻenehana.
  • E hōʻoia i ka ʻike pololei ʻana i ka ʻōlelo, a i ʻole e hoʻonohonoho lima.
  • No ka unuhi, e hōʻoia ʻaʻole ʻoe i hoʻohana iā turbo.
  • E hoʻohui i nā inoa mea ʻōlelo me kahi mea ʻē aʻe inā pono.
  • E loaʻa ka ʻae ma mua o ka hoʻopaʻa a hoʻololi ʻana i nā kānaka ʻē aʻe.

Nā nīnau pinepine

Manuahi anei ʻo Whisper AI?

ʻAe, manuahi ka hoʻoiho a holo ʻana iā OpenAI Whisper. Laikini ʻo OpenAI i ka code a me nā weights ma lalo o MIT, no laila hiki ke hoʻohana me ka uku ʻole iā OpenAI. Na ʻoe ke kamepiula, a pono ka model large i 10 GB GPU memory paha. Uku nā koho hoʻokipa: uku ka API a OpenAI no ka hoʻohana, a hoʻonohonoho nā polokalamu i kā lākou mau hoʻolālā. ʻAʻole hoʻi ka “manuahi” he maʻalahi; pono ʻo Python, FFmpeg a me ka kamaʻāina i ka command line.

Hiki iā Whisper ke ʻike i nā mea ʻōlelo ʻokoʻa?

ʻAʻole ma kāna iho. Kākau ʻo Whisper i ka mea i ʻōlelo ʻia akā ʻaʻole inoa i ka mea nāna i ʻōlelo. Wahi a ke kāleka OpenAI ʻaʻole i loiloi paʻa ʻia ka speaker diarization. No nā inoa, hoʻohui maʻamau nā kānaka iā Whisper me kahi mea diarization ʻē aʻe, a i ʻole hoʻohana i kahi polokalamu me ka ʻike mea ʻōlelo. No ka laʻana, loaʻa iā whispertranscribe.ai kahi koho hiki ke hoʻā ma mua o ka hoʻololi. E nānā mau i nā inoa ma nā leo e like ke kani.

Hiki iā Whisper ke hoʻololi i ka manawa maoli?

ʻAʻole me ka hana hou ʻole. Hana ʻia ʻo Whisper no nā leo i hoʻopaʻa ʻia a hana i nā puka 30 kekona; ʻōlelo ke kāleka OpenAI ʻaʻole hiki ke hoʻohana i ka manawa maoli me ka hana hou ʻole. Ua kūkulu nā mea hoʻomohala i nā mea kokoke i ka manawa maoli ma ka hoʻokomo ʻana i nā ʻāpana pōkole i ko lākou hiki ʻana mai. Inā pono nā live captions, e ʻimi i ka mea me ka mode ola a manaʻo e pono ka hoʻoponopono hou aʻe ma mua o kahi transcript faila piha.

ʻEhia ʻōlelo e kākoʻo ai ʻo Whisper?

ʻIke ʻo Whisper i 99 ʻōlelo paha, a papa inoa ka tokenizer i kēia manawa i 100 language codes ma hope o ka hoʻohui Cantonese ma large-v3. ʻOkoʻa nui ka maikaʻi. Wahi a OpenAI he hopena ikaika ma kahi o 10 ʻōlelo, a ʻike ka pepa he ma lalo o 1,000 hola aʻo ko ka hapa nui. Ma mua o ka hilinaʻi ʻana i kahi ʻōlelo kakaʻikahi, e hoʻāʻo i kahi ʻāpana pōkole. ʻO nā helu huahana, e like me ka 130+ ʻōlelo o whispertranscribe.ai, pili i ke kākoʻo o ia huahana, ʻaʻole ka model open-source.

Hoʻokahi anei ʻo whispertranscribe.ai a me OpenAI Whisper?

ʻAʻole. He model open-source ʻo OpenAI Whisper e hoʻoiho a holo ai ʻoe iho. He lawelawe transcription kaʻawale ma ka polokalamu kele pūnaewele ʻo whispertranscribe.ai; wehewehe ka ʻaoʻao mua ua hoʻohana ia iā Whisper a me AI hou. Hoʻohui ia i nā mea ʻaʻole i loko o ka model kumu, e like me ka hoʻouka a hiki i 5GB, nā loulou YouTube a Google Drive, ka ʻike mea ʻōlelo, a me nā puka DOCX, PDF, SRT a VTT. He mau huahana kaʻawale nā polokalamu ʻē aʻe me “Whisper” ma ka inoa.

No ke aha e kākau ai ʻo Whisper i nā huaʻōlelo i ʻōlelo ʻole ʻia?

No ka mea wānana ia i ka kikokikona e like me kahi language model. Wahi a ke kāleka OpenAI, hiki i ke aʻo ʻana ma nā ʻike pūnaewele nui a walaʻau ke hoʻohua i nā huaʻōlelo ʻaʻole i loko o ka leo. Wahi a OpenAI, hui ka model i ka wānana huaʻōlelo aʻe me ka hoʻololi ʻana i ka leo. Ua ʻike kahi noiʻi 2024 i nā ʻōlelo piha i haku ʻia ma 1% paha o nā transcripts i nānā ʻia, ʻoi pinepine ma nā hoʻomaha lōʻihi. E nānā i nā hoʻomaha, nā wahi mālie a me nā mea koʻikoʻi me ka leo.


Makemake i ka transcript Whisper me ka hoʻonohonoho ʻole? E hoʻouka i ka faila a hoʻopili paha i ka loulou ma whispertranscribe.ai a loaʻa ka kikokikona ma 130+ ʻōlelo me DOCX, PDF, SRT a VTT.

Nā kumu

Makemake ʻoe e hoʻolilo i kāu mau hoʻopaʻa leo i kikokikona?

E hoʻohana iā whispertranscribe.ai e kākau i ka leo o nā waihona leo a me nā wikiō.

Hoʻomaka i ke kākau leo ↗