Pehea ka pololei o Whisper AI? WER, nā palena a me nā ʻōlelo kōkua (2026)

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

Pehea ka pololei o Whisper? Ma ka ʻōlelo Pelekānia maʻemaʻe i heluhelu ʻia, hōʻike ka pepa a OpenAI i ka word error rate (WER) 2.5% no Whisper large-v2, ʻo ia hoʻi he 97 a 98 paha huaʻōlelo pololei mai 100. ʻOi aku ka hewa ma nā leo paʻakikī: 13.6% ma nā kelepona, 24.9% ma kahi pāʻina ʻaina walaʻau, a me 39.9% ma kahi hālāwai i hoʻopaʻa ʻia me hoʻokahi microphone mamao.

ʻO Whisper ke kumu hoʻohālike ʻike leo open-source a OpenAI a hoʻohana ʻia e nā polokalamu hoʻololi leo he nui, me kā mākou. Pili kēia alakaʻi i ka pololei wale nō: ka mea a WER e ana ai, nā hopena ma ke ʻano leo a me ka ʻōlelo, nā hemahema o Whisper, a me ka hoʻāʻo ʻana iā ʻoe iho. No ke kumu hoʻohālike, e ʻike i he aha ʻo Whisper AI.

Mai ka pepa Whisper ka hapa nui o nā helu (Radford a me nā hoa, 2022), i hoʻāʻo i nā model a pau ma ke ʻano “zero-shot” me ka ʻole o ke aʻo hou ma nā ʻike hoʻāʻo. Hōʻike ka kāleka model large-v3 hou aʻe i 10% a 20% ka emi o nā hewa ma mua o large-v2 ma nā ʻōlelo he nui, akā ʻaʻohe papa piha no kēlā me kēia dataset. Hoʻopuka ʻo whispertranscribe.ai i kēia ʻatikala. ʻAʻole mākou i hana i kahi hoʻāʻo hoʻohālikelike no kēia ʻatikala, a ʻaʻole nā helu ma ʻaneʻi nā hopena o kā mākou hoʻāʻo ponoʻī.

Ka pololei o Whisper i ka nānā pōkole

ʻAʻole lawa hoʻokahi dataset e pane ai, no ka mea hoʻololi ka leo i ka helu ma mua o nā mea ʻē aʻe. Hōʻike ka papa i nā hopena a OpenAI no large-v2 me beam search a me temperature fallback (papa 9 o ka pākuʻi), ma hope o ka normalizer kikokikona a OpenAI.

ʻAno leo Dataset ma ka pepa Whisper large-v2 WER Māhele huaʻōlelo pololei ma kahi kokoke
Puke leo, heluhelu maʻemaʻe LibriSpeech test-clean 2.5% 97.5 paha mai 100
Nūhou i heluhelu ʻia WSJ 2.6% 97.4 paha mai 100
Nā haʻiʻōlelo TED TED-LIUM 3 3.7% 96 paha mai 100
Puke leo, nā mea ʻōlelo paʻakikī aʻe LibriSpeech test-other 4.9% 95 paha mai 100
Nā mea manawaleʻa me nā microphone home Common Voice 5.1 8.2% 92 paha mai 100
Nā kamaʻilio kelepona Switchboard 13.6% 86 paha mai 100
Nā nīnauele sociolinguistic lōʻihi CORAAL 14.2% 86 paha mai 100
Nā kelepona ʻohana a hoaaloha CallHome 16.4% 84 paha mai 100
Nā hālāwai, microphone headset no kēlā me kēia kanaka AMI IHM 17.4% 83 paha mai 100
Nā pāʻina ʻaina, microphone mamao CHiME-6 24.9% 75 paha mai 100
Nā hālāwai, hoʻokahi microphone papaʻaina mamao AMI SDM1 39.9% 60 paha mai 100

ʻO ka “māhele huaʻōlelo pololei” he 100 i unuhi ʻia ʻo WER, he helu kokoke wale nō, no ka mea helu pū ʻo WER i nā huaʻōlelo i hoʻohui ʻia.

ʻIke pinepine ʻia ka 2.7%: ʻo ia ka model like me ka greedy decoding maʻalahi aʻe (papa 2 o ka pepa). ʻO nā laina AMI he mau hālāwai like i hoʻopaʻa ʻia ma nā ʻano ʻelua. Ke hoʻoneʻe ʻia ka microphone mai ke poʻo o kēlā me kēia mea ʻōlelo i ke kikowaena o ka papaʻaina, ʻoi aku ka helu hewa ma mua o ka pālua.

He aha ka word error rate (WER), a pehea e helu ai?

ʻO WER ka nui o nā hewa i hoʻohālikelike ʻia me kahi transcript kuhikuhi i nānā ʻia e ke kanaka. E helu i nā hewa ʻekolu a puʻunaue i ka nui huaʻōlelo o ka transcript kuhikuhi:

WER = (nā pani + nā mea i waiho ʻia + nā mea i hoʻohui ʻia) ÷ nā huaʻōlelo o ke kuhikuhi

  • Pani: he huaʻōlelo hewa ma kahi o ka huaʻōlelo pololei (“lunch” ma kahi o “launch” ma ka ʻōlelo Pelekānia).
  • Waiho: he huaʻōlelo i ʻōlelo ʻia akā nalowale i ka transcript.
  • Hoʻohui: he huaʻōlelo ma ka transcript i ʻōlelo ʻole ʻia e kekahi.

Eia ka laʻana Pelekānia i hana ʻia e nā mea kākau o ka ʻatikala kumu, i helu ʻia me ka waihona open-source jiwer:

Laina Kikokikona
Kuhikuhi (22 huaʻōlelo) we moved the product launch to march ninth because the vendor missed two deadlines and doctor patel wants a new test plan
Transcript we moved the product lunch to march ninth because vendor missed two deadlines and doctor battle wants a new test plan today
Nā hewa 2 pani (lunch, battle), 1 waiho (the), 1 hoʻohui (today)
WER 4 ÷ 22 = 18.2%

ʻEkolu mea e kānalua ai nā kānaka e pili ana iā WER:

  1. Hiki ke ʻoi ma mua o 100%. ʻAʻohe palena o nā huaʻōlelo i hoʻohui ʻia, a ʻoi aku kekahi mau ʻōlelo me ka ʻike aʻo liʻiliʻi ma mua o 100% ma nā hopena Fleurs o ka pepa.
  2. Hoʻololi ka normalization i ka helu. Helu ʻia “Dr.” a me “doctor” he hewa inā ʻaʻole hoʻomaʻemaʻe like ʻia nā kikokikona ʻelua. Wahi a OpenAI, ua hōʻemi kāna normalizer i WER a hiki i 50% ma kekahi datasets, ma ka wehe nui ʻana i nā ʻokoʻa ʻano kākau.
  3. Hoʻohana kekahi mau ʻōlelo i ka character error rate (CER). No ka ʻōlelo Kina, Kepanī, Thai, Lao a me Burmese, kahi e hoʻokaʻawale ʻole ʻia ai nā huaʻōlelo me nā hakahaka, helu ka pepa i nā huapalapala.

ʻO ka ʻōlelo no ka pololei me ka inoa ʻole o ka dataset a me ke ʻano helu, ʻaʻole ia he ana, ʻo wai ka mea e ʻōlelo ai, me mākou pū.

Pehea ka pololei o Whisper ma ka ʻōlelo Pelekānia?

Pololei loa ʻo Whisper ma ka Pelekānia maʻemaʻe a hiki nō ke hoʻohana ma ka leo paʻakikī. Ma ka pepa, he 55.2% ka emi o kāna mau hewa ma nā datasets Pelekānia 12 ʻē aʻe ma mua o kahi model wav2vec 2.0 i aʻo ʻia me ke alakaʻi, me ka helu LibriSpeech aneane like.

Nā leo lōʻihi

Heluhelu ʻo Whisper i ka leo ma nā puka 30 kekona, no laila hoʻāʻo ʻokoʻa ka pepa i nā faila lōʻihi (papa 16 o ka pākuʻi, large-v2):

Dataset lōʻihi Ka ʻike i loko WER
TED-LIUM 3 11 haʻiʻōlelo TED piha 3.5%
Meanwhile 64 ʻāpana kīwī pō, he nui nā huaʻōlelo ʻano ʻē 5.1%
Kincaid46 46 leo like ʻole mai ka hoʻohālikelike 2018 8.8%
Earnings-21 Nā kelepona hōʻike kālā hui 9.7%
Rev16 16 ʻāpana podcast 11.3%
Earnings-22 Nā kelepona hōʻike kālā hou me nā leo ʻokoʻa he nui 12.6%
CORAAL 231 nīnauele piha ma ka African American English 19.6%

Whisper a me nā mea hoʻololi leo kanaka

Ua noi ʻo OpenAI i ʻehā lawelawe kanaka a hoʻokahi lawelawe kōkua kamepiula e hoʻololi i 25 leo Kincaid46. Ua ʻoi aku ka lawelawe kōkua kamepiula ma mua o Whisper ma 1.15 percentage points, a ʻoi iki wale nō nā lawelawe kanaka. Wahi a ka pepa, ʻaʻole maikaʻi loa akā kokoke loa i ka pololei o ke kanaka.

He mea nui ka nui o ka model

Loaʻa ʻo Whisper ma nā nui he nui, a ʻaʻole haʻi mau nā polokalamu i ka model a lākou e hoʻohana ai (papa 9 o ka pākuʻi):

Model Nā parameters LibriSpeech clean Common Voice 5.1 Switchboard CHiME-6
tiny 39M 6.7% 26.1% 18.3% 48.0%
base 74M 4.9% 19.4% 15.6% 40.0%
small 244M 3.3% 12.8% 13.3% 30.9%
medium 769M 2.7% 9.7% 13.2% 27.6%
large-v2 1550M 2.5% 8.2% 13.6% 24.9%

Nui aʻe ka ʻokoʻa ke paʻakikī ka leo: 4 points paha ma nā puke leo maʻemaʻe, 23 points ma ka pāʻina. ʻO turbo hou aʻe he large-v3 i hoʻomākaukau ʻia no ka wikiwiki; ʻae ia i ka emi liʻiliʻi o ka pololei e like me ka GitHub README a OpenAI.

Pehea ka pololei o Whisper ma nā ʻōlelo ʻē aʻe?

Ma waho o ka Pelekānia, pili ka pololei i ka nui o ka leo aʻo i loaʻa iā Whisper ma ia ʻōlelo. Ua loaʻa i ka pepa he squared correlation 0.83 ma waena o ka logarithm o nā hola aʻo a me WER, a kuhi e emi ka WER i ka hapalua no kēlā me kēia hoʻonui 16 pā o ka ʻikepili. Ua loaʻa i ka hapa nui o nā ʻōlelo 96 ʻaʻole Pelekānia ma lalo o 1,000 hola.

Nā hopena Fleurs i koho ʻia (nā ʻōlelo i heluhelu ʻia, large-v2, papa 13 o ka pākuʻi):

ʻŌlelo Fleurs WER Kaha
Paniolo 3.0% Haʻahaʻa loa ma ka papa
Pelekānia 4.2%
Pukikī 4.3%
Kelemānia 4.5%
Kepanī 5.3% Character error rate
Lūkini 5.6%
Palani 8.3%
Kōlea 14.3% Hoʻomaopopo ʻia he hopena ʻokoʻa
Kina 14.7% Character error rate
ʻAlapia 16.0%
Hindi 21.5%
Hebera 27.1% Hoʻomaopopo ʻia he hopena ʻokoʻa
Swahili 39.3%
Telugu 99.0% Hoʻomaopopo ʻia he hopena ʻokoʻa

Hōʻike ka pepa he ʻoi aku ka maikaʻi ʻole o Hebera, Telugu, Kina a me Kōlea ma mua o ka mea i kuhi ʻia mai kā lākou ʻike aʻo, no ka mamao ʻōlelo paha, tokenizer kūpono ʻole i kēlā mau palapala, a i ʻole ka maikaʻi ʻikepili. He ʻōlelo heluhelu ʻo Fleurs, no laila e manaʻo i nā helu hewa kiʻekiʻe aʻe ma ke kamaʻilio. Aia ma ka README kekahi pakuhi hou no large-v3 a me large-v2 ma Common Voice 15 a me Fleurs.

He mau hopena hoʻoikaika nā hopena Paniolo inā hoʻololi ʻoe i ka leo Paniolo i kikokikona. No nā ʻōlelo me ka ʻike aʻo liʻiliʻi, e hoʻāʻo i kahi ʻāpana pōkole ma mua o ka papahana holoʻokoʻa.

He aha ka mea e hoʻopilikia i ka pololei o Whisper?

ʻO nā kūlana hoʻopaʻa leo ka mea nui loa. Nā hōʻike mai ka pepa a me ke kāleka model:

Mea Ka mea a ka hōʻike e hōʻike ai Ka hana e kōkua ai
Walaʻau hope Ma ka walaʻau hale inu ma lalo o 10 dB signal-to-noise ratio, ua ʻoi ʻo Whisper ma luna o nā model hoʻohālikelike 14, akā ua emi nā model a pau i ka piʻi ʻana o ka walaʻau Hoʻopaʻa i loko o kahi lumi mālie, mamao i nā peʻahi a me nā mele
Mamao microphone Nā hālāwai AMI like: 17.4% me nā headset, 39.9% me hoʻokahi microphone papaʻaina E hoʻokokoke i ka microphone i ka mea ʻōlelo
Kamaʻilio a me nā leo e uhi ana 13.6% a 24.9% no nā kelepona a me nā pāʻina, me 2.5% no ka ʻōlelo heluhelu Hoʻokahi mea ʻōlelo i ka manawa; hoʻokahi microphone no kēlā me kēia
Nā leo kūloko a me nā dialect Hōʻike ke kāleka model i ka pololei like ʻole; 19.6% nā nīnauele CORAAL lōʻihi E hoʻāʻo me nā mea ʻōlelo e like me kāu mau leo maoli
Nā huaʻōlelo kakaʻikahi a kūikawā Maikaʻi ma ka Meanwhile me nā huaʻōlelo kūikawā (5.1%), akā mau nā hewa inoa a me nā huaʻōlelo ʻenehana E nānā i nā inoa a me nā huaʻōlelo me kahi papa inoa
Ka nui ʻikepili o ka ʻōlelo ʻO Fleurs WER mai 3.0% (Paniolo) a i 99.0% (Telugu) E koho i ka ʻōlelo iā ʻoe iho; e hoʻolālā i ka nānā hou no nā ʻōlelo me ka ʻikepili liʻiliʻi
Ka nui model 48.0% ʻo tiny ma CHiME-6, 24.9% ʻo large-v2 E nīnau i ka model a ka mea hana e hoʻohana ai, a i ʻole e hoʻohana i ka nui loa i loaʻa
Ke ʻano faila a me ka sample rate Hoʻololi hou ka code Whisper i nā leo a pau i 16 kHz mono ʻAʻole hoʻohui ka 48 kHz WAV i ka maikaʻi ma mua o kahi faila 16 kHz maʻemaʻe

ʻAʻole hoʻihoʻi ka hoʻololi ʻana i MP3 i WAV nui i nā kikoʻī i nalowale. ʻOi aku ka leo maʻemaʻe me ka bitrate waena ma mua o ka leo walaʻau me ka bitrate kiʻekiʻe.

Nā pilikia Whisper i ʻike ʻia: hallucinations, nā loop a me nā manawa

Kākau ʻo OpenAI i kēlā me kēia pilikia; hiki ke koʻikoʻi lākou ma mua o kā lākou māhele ma WER.

Kikokikona i haku ʻia

Hiki iā Whisper ke kākau i nā huaʻōlelo i ʻōlelo ʻole ʻia. Wahi a ke kāleka model hiki i nā wānana ke loaʻa nā kikokikona i ʻōlelo ʻole ʻia i ka leo, no ka mea wānana pū ka model i nā huaʻōlelo e hiki mai ana. Hōʻike ka ʻāpana palena o ka pepa i ka hallucination piha, he hopena pili ʻole i ka leo.

Ua hoʻohana ka noiʻi kūʻokoʻa Careless Whisper (Koenecke a me nā hoa, 2024) i 13,140 ʻāpana leo ma ka Whisper API a OpenAI i ʻApelila a me Mei 2023. Ma kahi o 1% o nā transcripts (1.4% ma ka awelika) i loaʻa nā huaʻōlelo a ʻōlelo piha i haku ʻia, a he 38% o ia mau mea i loaʻa ka ʻike ʻino e like me ka hana ʻino a i ʻole ka hoʻopili wahaheʻe. Ua ʻoi aʻe nā mea ʻōlelo me nā hoʻomaha lōʻihi, ma ʻaneʻi ka poʻe me ka aphasia. Ua loli ka API mai ia manawa, no laila e nānā i kēia he ana 2023.

Nā loop hou a me nā huaʻōlelo nalowale

Wahi a ke kāleka model he ʻano ko Whisper e hana i ka kikokikona hou; hoʻemi ka beam search a me temperature scheduling i kēia, akā ʻaʻole piha. Papa inoa ka pepa i ka waiho ʻana i nā huaʻōlelo mua a hope paha o kahi ʻāpana. Ua hōʻemi nā hoʻoponopono decoding i ka WER awelika o ka leo lōʻihi mai 11.0% a i 10.0%: ʻoi aku ka maikaʻi, ʻaʻole i hoʻopau ʻia ka pilikia.

Ka hāmau, ke mele a me nā huaʻōlelo hoʻopiha

ʻIke pinepine ʻia ka kikokikona i haku ʻia ma nā hāmau lōʻihi a me ke mele. Ua ʻike ka pepa ʻaʻole hiki ke hilinaʻi i ka hōʻailona “ʻaʻohe ʻōlelo” a Whisper wale nō, a aia i ka code open-source kahi koho e lele i nā hāmau lōʻihi inā ʻike ʻia kahi hallucination paha.

Ka neʻe o nā manawa a me nā faila lōʻihi

Hana ʻo Whisper i ka leo lōʻihi ma nā puka 30 kekona, me ka hoʻohana i kāna mau manawa i wānana ʻia e kau i ka puka aʻe. Wahi a ka pepa hiki i ka hewa ma kekahi puka ke neʻe i ka puka aʻe. Kaha ʻia nā manawa huaʻōlelo “experimental” ma ka code. No nā subtitles, e nānā i nā manawa ma kahi kokoke i nā hoʻomaha.

I ka wā hoʻomohala, ʻike pū ʻo OpenAI ua koho ʻo Whisper i nā inoa mea ʻōlelo a aneane hewa mau. Ua aʻo hou ʻia nā model i hoʻokuʻu ʻia e hōʻemi i kēia, akā e nānā i kēlā me kēia inoa ma ka kikokikona.

10 ala e hoʻomaikaʻi ai i ka pololei o Whisper

Hiki mai ka nui o nā hoʻomaikaʻi ma mua o ka hoʻomaka ʻana i ka hoʻololi. Mai ka hopena nui a i ka liʻiliʻi ma kahi kokoke:

  1. E hoʻokokoke i ka microphone. ʻO ka mamao wale nō i hoʻonui i ka helu hewa ma mua o ka pālua ma nā hālāwai AMI.
  2. E hoʻemi i ka walaʻau ma ke kumu. E pani i nā puka makani, hoʻopau i nā peʻahi a me ke mele.
  3. E pale i nā leo e uhi ana. Hoʻokahi kanaka i ka manawa ka mea ʻoi aku ma mua o kekahi hoʻonohonoho.
  4. E ʻoki i nā hāmau lōʻihi a me nā hoʻomaka mele, kahi e ʻike pinepine ʻia ai ka kikokikona i haku ʻia. E ʻoki ma ka hoʻomaha, ʻaʻole ma waena o kahi ʻōlelo.
  5. E koho i ka ʻōlelo iā ʻoe iho. Ua ʻike pololei ʻo Whisper i ka ʻōlelo 80.3% o ka manawa ma nā ʻōlelo Fleurs 82, no laila hiki ke hewa ka ʻike ʻakomi ma nā ʻāpana pōkole a me nā leo kūloko.
  6. E hoʻohana i ka model nui loa i loaʻa, ʻoiai no nā leo walaʻau.
  7. E hāʻawi i nā inoa a me nā huaʻōlelo inā mālama ʻoe i ka ʻōnaehana iā ʻoe iho. ʻAe ʻo Whisper open-source i --initial_prompt me nā ʻano kākau i manaʻo ʻia.
  8. E hoʻouka i ka faila kumu. Hoʻopaʻa hou ma nā speaker a compress hou paha i nā manawa he nui e nalowale ai nā kikoʻī.
  9. E nānā i nā huaʻōlelo. E huli i kēlā me kēia inoa, helu, lā a me ka pōkole, a hōʻoia me ka leo.
  10. E nānā ʻelua manawa. E hoʻoponopono mua i nā hewa ma ka wikiwiki maʻamau. A laila e ʻimi i nā hōʻailona hallucination: nā laina hou, nā ʻōlelo i ka hāmau, nā hoʻopau e like me “Mahalo no ka nānā ʻana.” No nā hana kānāwai, olakino a paʻi ʻia paha, e nānā ke kanaka i nā mea a pau.

Pehea e ana ai i ka pololei o Whisper ma kāu leo ponoʻī

ʻO ke ana maikaʻi loa no kāu hana he hoʻāʻo 30 minuke ma kāu mau leo. No ʻoe kēia mau ʻanuʻu e hana ai; ʻaʻole ia he hoʻāʻo a whispertranscribe.ai.

  1. E koho i ʻekolu ʻāpana 2 minuke e like me kāu hana maoli: hoʻokahi maʻalahi, hoʻokahi maʻamau a hoʻokahi paʻakikī.
  2. E kākau lima i kahi transcript kuhikuhi, huaʻōlelo ma ka huaʻōlelo. E hoʻoholo mua i nā lula: mālama a waiho paha i ka “um”, kākau helu me nā huahelu a me nā huaʻōlelo paha.
  3. E hoʻololi i kēlā me kēia ʻāpana me ka mea hana āu e hoʻāʻo nei a mālama i ka kikokikona maʻalahi.
  4. E normalize like i nā kikokikona ʻelua. Hele pū ka normalizer a OpenAI me openai-whisper.
  5. E helu iā WER me jiwer:
# pip install jiwer openai-whisper
import jiwer
from whisper.normalizers import EnglishTextNormalizer

norm = EnglishTextNormalizer()
ref = norm(open("reference.txt", encoding="utf-8").read())
hyp = norm(open("transcript.txt", encoding="utf-8").read())

out = jiwer.process_words(ref, hyp)
print(f"WER {out.wer:.1%}  S={out.substitutions} D={out.deletions} I={out.insertions}")
  1. E helu kaʻawale i nā hewa nui. Hāʻawi nā huaʻōlelo hoʻopiha ʻelima hewa a me nā inoa mea kūʻai ʻelima hewa i ka WER like, akā ʻaʻole ka pilikia like.
  2. E hoʻopaʻa i nā hopena no ka hoʻohālikelike i nā mea hana a hoʻonohonoho paha:
ʻĀpana Kūlana Nā huaʻōlelo kuhikuhi S D I WER Nā hewa nui
ʻĀpana A Lumi mālie, hoʻokahi mea ʻōlelo 310
ʻĀpana B Kelepona wikiō, ʻelua mea ʻōlelo 285
ʻĀpana C Hale kope, mea ʻōlelo me ka leo kūloko 296

He mau helu laʻana nā huaʻōlelo. No ka ʻanuʻu 4: i ka wā i hoʻohana ai nā mea kākau o ka ʻatikala Pelekānia kumu i ka normalizer Pelekānia a OpenAI ma kā lākou ʻōlelo i haku ai, ua hoʻohālikelike ia iā “Dr.” me “doctor” a me “two” me “2”, akā helu nō ia iā “9th” a me “ninth” he hewa. E heluhelu i nā ʻokoʻa i hōʻike ʻia ma mua o ka hilinaʻi ʻana i ka helu.

Pehea ka pololei o whispertranscribe.ai?

He mea hoʻololi leo AI ma ka pūnaewele ʻo whispertranscribe.ai i hoʻohana iā Whisper, e kākoʻo ana i 130+ ʻōlelo, nā hoʻouka a hiki i 5GB a me 10 hola. Hiki ka pololei i 99% ma ka leo maʻemaʻe; hoʻemi ka walaʻau hope, nā leo ʻokoʻa ikaika a me nā leo e uhi ana, e like me nā hoʻāʻo Whisper ma luna. ʻAʻole mākou e ʻōlelo ua like a ʻoi aku paha mākou ma mua o nā helu o ka pepa, a ʻaʻole mākou i paʻi i kā mākou hoʻāʻo hoʻohālikelike; e hoʻāʻo i kāu mau ʻāpana me nā ʻanuʻu ma luna.

Hiki iā ʻoe ke hoʻomaka manuahi me ka ʻole o ka moʻokāki. Hoʻomaka nā hoʻolālā uku ma $9.99/mahina ke uku makahiki ʻia; ʻike i ke kumukūʻai.

Nā nīnau pinepine

He 99% anei ka pololei o Whisper?

ʻAʻole ma nā kūlana a pau. ʻO ka hopena maikaʻi loa i paʻi ʻia he 2.5% WER ma LibriSpeech test-clean, nā puke leo i heluhelu maʻemaʻe ʻia, ma kahi o 97.5% huaʻōlelo pololei. Ma nā kelepona, hālāwai a me nā lumi walaʻau, ʻo nā helu a OpenAI ma kahi o 13% a 40% WER. He manaʻo ko ka “99%” inā hōʻike ia i ka leo i ana ʻia. Pono e loiloi i nā mea hana ma Whisper, me whispertranscribe.ai, ma kāu mau leo ponoʻī, ʻaʻole ma ka pākēneka nui wale nō.

ʻO wai ka model Whisper pololei loa?

ʻO nā model nui ka pololei loa. Ma ka pepa OpenAI, loaʻa iā large-v2 nā helu hewa haʻahaʻa loa ma ka hapa nui o nā datasets, a hōʻike ke kāleka large-v3 i 10% a 20% ka emi o nā hewa ma nā ʻōlelo he nui. He large-v3 i hoʻomākaukau ʻia ʻo turbo e nalowale liʻiliʻi ai ka pololei no ka wikiwiki nui e like me OpenAI. ʻOi aku ka emi o tiny ma nā leo paʻakikī: 48.0% WER ma ka pāʻina walaʻau, me 24.9% no large-v2.

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

Hiki i Whisper ke hallucinate no ka mea ua aʻo ʻia e wānana i ka kikokikona e hiki mai ana a me ka hoʻololi leo; ʻōlelo ke kāleka OpenAI hiki iā ia ke hoʻopuka i ka kikokikona ʻaʻole ma ka leo. ʻIke pinepine ʻia i nā hāmau lōʻihi, mele a walaʻau paha. Ua ʻike kekahi noiʻi kūʻokoʻa 2023 i nā huaʻōlelo i haku ʻia ma kahi o 1% o nā transcripts, a hoʻonui nā hoʻomaha i ka manawa e kū ai. Hoʻemi ka ʻoki hāmau, ke koho ʻōlelo a me ka nānā i nā ʻāpana mālie i ka pilikia.

Pehea ka pololei o Whisper me nā leo kūloko?

Hana ʻo Whisper me nā leo kūloko ma mua o nā ʻōnaehana kahiko he nui, akā ʻaʻole like ka pololei. Wahi a ke kāleka OpenAI ʻokoʻa ka pololei ma nā leo kūloko a me nā dialect o ka ʻōlelo hoʻokahi. Ma ka pepa, he 19.6% WER nā nīnauele lōʻihi African American English (CORAAL), a he 3.5% nā haʻiʻōlelo TED piha, ʻoiai ʻokoʻa pū nā kūlana hoʻopaʻa. ʻO ke ala hilinaʻi ka hoʻāʻo ʻana i kahi ʻāpana me nā mea ʻōlelo e like me kāu mau leo maoli, a hoʻolālā i ka nānā hou inā ikaika nā leo kūloko.

Like anei ka pololei o Whisper me ke kanaka hoʻololi leo?

No ka Pelekānia, kokoke ʻo Whisper akā ʻaʻole like me nā mea maikaʻi loa. Ua hoʻohālikelike ʻo OpenAI me ʻelima lawelawe ʻoihana ma 25 leo. He 1.15 points ka emi o ka WER o kahi lawelawe kōkua kamepiula ma mua o Whisper, a ʻoi iki nā lawelawe kanaka ʻehā. Hiki i nā kānaka ke mālama maikaʻi aʻe i nā inoa, nā leo e uhi ana a me ka ʻōlelo maopopo ʻole. No nā mea e ʻōlelo ʻia, waiho ʻia a paʻi ʻia paha, pono ke kanaka e nānā i ka transcript me ka leo.

Pehea ka pololei o Whisper ma nā ʻōlelo ʻē aʻe?

ʻOkoʻa nui ma ka ʻōlelo. Ma Fleurs, loaʻa iā large-v2 3.0% WER ma Paniolo, 4.5% ma Kelemānia a me 8.3% ma Palani, akā 27.1% ma Hebera a me 99.0% ma Telugu. Ana ʻia ka Kina a me Kepanī me character error rate: 14.7% a me 5.3%. Ua ʻike ka pepa pili ka pololei i ka nui o ka leo aʻo no kēlā me kēia ʻōlelo. No nā ʻōlelo me ka ʻikepili liʻiliʻi, e hoʻāʻo mua i kahi ʻāpana pōkole a hāʻawi manawa hou no ka hoʻoponopono.

E hoʻāʻo i ka pololei ma kāu leo ponoʻī

Hōʻike nā hoʻāʻo hoʻohālikelike i nā leo o nā kānaka ʻē aʻe. E hoʻouka i kahi ʻāpana i whispertranscribe.ai ma kekahi o nā ʻōlelo 130+ a helu me nā ʻanuʻu ma kēia alakaʻi. E hoʻomaka manuahi me ka ʻole o ka moʻokāki, a i ʻole hoʻohālikelike i nā hoʻolālā.

Nā ʻatikala pili: He aha ʻo Whisper AI? · Nā polokalamu hoʻololi leo manuahi maikaʻi loa · Nā koho TurboScribe maikaʻi loa · Pehea e hoʻololi ai i kahi wikiō YouTube · He aha ka faila SRT? · Nā huaʻōlelo paʻa a me nā mea hiki ke hoʻololi

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 ↗