InoteveraGaidhi rinotevera
Local Outlier Factor
Tekinoroji
Nhungamiro yehunyanzvi
Faster-Whisper uye whisper.cpp inguva dzenharaunda dzekumhanyisa mamodheru ekuziva Whisper matauriro ane mashandiro akasiyana uye kushambadzira.
Teerera kugwaro iriAI-yakarondedzerwa · ~3 min
Vanogona kuita kuti zvinyorwa zvemuno zvishande pane zvinotsigirwa zvehardware, asi kumhanya, ndangariro, hunhu hwemutauro, uye kurongeka zvinoenderana nemuenzaniso, quantization, uye nguva yekumhanyisa marongero.
Whisper imhuri yemhando dzekuziva kutaura dzakadzidziswa mabasa akadai sekunyora nekushandura nemitauro yakawanda. Faster-Whisper chiitiko chinomhanyisa Whisper modhi kuburikidza neCTranslate2, nepo whisper.cpp iri C uye C++ chiteshi chakagadzirirwa kumhanya pane dzakasiyana siyana dzepuratifomu. Ose ari maviri anovavarira kuita kuti inferensi ishande zvakanyanya kana kutakurika pane yakatwasuka referensi setup, asi ivo vanozviita kuburikidza akasiyana software stacks. Kuita kwenzvimbo kunogona kuderedza kudiwa kwekurodha odhiyo kune yekunze sevhisi uye kunogona kubatsira nekuita offline kugadzirisa. Zvinoreva zvakare kuti mushandisi anotarisira kurodha modhi, kuchengetedza, kutsamira panguva yekumhanya, uye kuenderana kwehardware. Iyo modhi faira inogona kutora yakakura dhisiki nzvimbo, uye inference kumhanya kunoenderana nemuenzaniso saizi, CPU kana GPU, inotsigirwa kukurumidza, batching, quantization, uye kureba kwekuteerera. Quantization inogona kuderedza ndangariro kana kuvandudza mabudiro, asi inogona kukanganisa kunyorwa kwemhando; kuyera tradeoff pane kufunga kuti haina basa. Iyo nguva dzekumhanya dzinogona kusiyana mukumisikidza, inotsigirwa modhi shanduko, mudziyo unodzosera kumashure, maitiro ekutepfenyura, uye decoding sarudzo. Feature rutsigiro inoshanduka, saka tarisa zvinyorwa zvazvino zveiyo chaiyo vhezheni uye hardware. Benchmark ine zvazvinoreva chete kana ichidoma iyo modhi, chaiyo, danda kana decoding marongero, nguva yekupinda, mudziyo, uye evaluation set. Chaiyo-nguva kuita pane kapfupi clip yakachena hakufanotaura kurekodha kurekodha nevatauri vanopindirana kana ruzha. Whisper inogadzira zvinyorwa zvichiri kuda kuongororwa. Mazita akakodzera, mazwi akasarudzika, matauriro, matauriro emumashure, uye vatauri vanopindirana zvinogona kukonzera kukanganisa. Chinyorwa chinofanira kuongororwa kana kukanganisa kuchikanganisa sarudzo, marekodhi, kana makotesheni. Izwi rekukanganisa mwero rinogona kuenzanisa hypotheses nemareferensi, asi kubatsira kwayo kunoenderana neyakajairwa kokorodzano uye mumiriri data data. Kumhanya munharaunda hakusi kwakavanzika kana kuenderana. Simbisa kwakachengetwa maodhiyo nemamodheru, angave marogi achichengeta zvirimo, ndiani anogona kuwana mafaera, uye inodiwa mvumo uye yekuchengeta zvinodiwa. Chengetedza modhi neshanduro dzenguva yekumhanya dzakapinirwa, uye sanganisira wongororo yevanhu kana kunyora kwacho kuchine basa.
Mutori wenhau anonyora mabvunzurudzo panzvimbo yebasa asingatumire zvakarekodhwa kune inotambirwa sevhisi, obva aongorora mazita nematemu ehunyanzvi.
Mugadziri anoisa mabhenji eFaster-Whisper paGPU inoenderana uye whisper.cpp palaptop CPU uchishandisa maredhiyo akafanana uye mhando dzemhando dzemhando.
Chikwata chinoziva kuvanzika chinotarisa marezenisi emodhi, hunhu hwenzvimbo yecache, uye telemetry isati yagadzirisa zvakarekodhwa.
Injiniya anoenzanisa quantized uye yakazara-chaiyo modhi akasiyana pane inomiririra yekuongorora seti, kusanganisira izwi kukanganisa mwero uye ndangariro kushandiswa.
Nguva dzekutaura dzemunharaunda dzicharamba dzichibatsirikana kubva kune zvinokurumidza kugadzirisa, kernels dzakagadziridzwa, uye madiki emuenzaniso anomiririra. Izvo zvinogona kuwedzera kushandiswa kwekunze pazvishandiso zvemunhu uye edge Hardware, nepo mapurojekiti akasiyana achishanduka pamitengo yakasiyana. Vashandisi vanofanirwa kutarisira kutsigirwa uye kuita kunosiyana nepuratifomu uye kuburitswa. Kuendesa kunoshanda kunochengeta kuenzanisa kumhanya, kurongeka, kuchengetedza modhi, bhatiri kana ndangariro bhajeti, zvidzoreso zvekuvanzika, uye ongororo yemunhu yezvikanganiso-zvinotapukira. Kuenderana kuyedzwa kunofanirwa kudzokororwa mushure menguva yekumhanya kana shanduko yehardware. Dzokorora mabhenji mushure mekusimudzira.
Tsamba yemahara
Nhatu dzakasimbiswa dzeAI nyaya mangwanani ega ega pakati pevhiki, dzakanyorwa neChirungu chakajeka. Yemahara zvachose, hapana ads.
Imwe email zuva rega rega revhiki. Bvisa kunyoresa kamwe chete. Hatimbotengese kana kugovera kero yako.
Zviedze iwe pachako
Mhinduro yepakarepo pamhinduro yega yega, uye chitupa chinogovaniswa chine ID inobatika kana wapedza kosi.
Tsigira dzidzo yemahara yeAI. AI Understanding ndeye 501(c)(3) isingabatsiri - hapana ads, hapana paywall, zvachose. Ita chipo
Faster-Whisper uye whisper.cpp inguva dzenharaunda dzekumhanyisa mamodheru ekuziva Whisper matauriro ane mashandiro akasiyana uye kushambadzira. Vanogona kuita kuti zvinyorwa zvemuno zvishande pane zvinotsigirwa zvehardware, asi kumhanya, ndangariro, hunhu hwemutauro, uye kurongeka zvinoenderana nemuenzaniso, quantization, uye nguva yekumhanyisa marongero.
Inokurumidza-Whisper inoshandisa Whisper inference uchishandisa CTranslate2.
Quantization inoshandura nhamba chaiyo uye inogona kukanganisa kubuda kwemhando.
Benchmark mamiriro anotaridza kuyerwa kwekubuda uye inofanirwa kuenzaniswa.
WER inofananidza mazwi anozivikanwa ane mareferensi, saka tokenization uye normalization zvibvumirano zvine basa.
Kugadziriswa kwenzvimbo kunoshandura kuyerera kwedata asi hakugadzirise zvese zvakavanzika uye zvekutonga zvinodiwa.
Ramba uchidzidza
Mamwe madhairekitori akasarudzirwa nyaya iyi
InoteveraGaidhi rinotevera
Local Outlier Factor
Tekinoroji