Amamodeli Angcono Kakhulu Okuqaphela Inkulumo Evulekile (ASR) ngo-2026: I-WER, Izilimi, Ukubambezeleka, kanye Nelayisensi Kuqhathaniswa

Ukuqashelwa kwenkulumo evulekile kuyekile ukuba yi-Whisper monoculture esikhathini esithile ezinyangeni eziyishumi nambili ezedlule. NgoMashi 2026 u-Cohere ukhiphe i-Transcribe, imodeli engu-2B Apache 2.0 ethathe phezulu kubhodi yabaphambili ye-Hugging Face Open ASR ngo-5.42% wesilinganiso samaphutha esimaphakathi. Emasontweni amahlanu kamuva i-IBM yathumela i-Granite Speech 4.1 2B ngo-5.33%. Kusukela lapho i-ARK-ASR-3B kanye ne-MOSS-Transcribe-preview-2B bathumele izinombolo eziphansi.

Phezulu kwalelo bhodi labaphambili manje kuhlukaniswe iphoyinti le-WER elingaphansi kwelilodwa. Lokho kunomphumela othize kunoma ubani okhetha imodeli: izinga aliselona uhlobo olunqumayo. Ilayisense, ukufakwa kolimi, ukwesekwa kokusakaza, kanye nezindleko ngehora ngalinye lomsindo kukhona. Le Roundup iqhathanisa inkundla kuzo zozine.

Okokuqala, inkinga ngenombolo yebhodi yabaphambili ecashunwa yiwo wonke umuntu

I-Open ASR Leaderboard avareji ayilona inani elingashintshi elilodwa, futhi amamodeli okwamanje asohlwini oluhlangene awazange athole amaphuzu ngendlela efanayo.

U-Cohere ongu-5.42% uyisilinganiso kuwo wonke amasethi okuhlolwa esiNgisi ayisishiyagalombili, okuhlanganisa i-TED-LIUM. Okubalulekile: AMI 8.13, Earnings-22 10.86, GigaSpeech 9.34, LibriSpeech clean 1.25, LibriSpeech other 2.37, SPGISpeech 3.08, TED-LIUM 2.49, VoxPopuli 4 to 5.8 exactly.

I-ARK-ASR-3B engu-5.04% iyisilinganiso sonkana Isikhombisa amasethi. I-TED-LIUM ayikho. Ikhadi le-MOSS-Transcribe-preview-2B lisho lokhu ngokucacile: I-TED-LIUM okwamanje ayiyona ingxenye yokugijima kwebhodi labaphambili ngakho ayifakiwe.

I-TED-LIUM ingenye yamasethi alula ku-suite, ngakho ukuyiyeka kuphakamisa isilinganiso. Bala futhi amaphuzu ka-Cohere ashicilelwe kudathasethi ngayinye kumasethi ayisikhombisa emibiko ye-ARK futhi u-Cohere ufike ku-5.84, hhayi ku-5.42. Yenza okufanayo ku-Granite Speech 4.1 2B futhi isuka ku-5.33 iye ku-5.65. Ngokwesisekelo sokufana nokufana nokuhola kwe-ARK kungu ezinkulu kunokuba izinombolo zesihloko zisho, hhayi ezincane – kodwa iphuzu liwukuthi awukwazi ukususa umfanekiso oshicilelwe komunye futhi uthole impendulo enengqondo.

Ezinye izixwayiso ezimbili zisekhasini elifanayo:

Ezinye izikolo zifakwe ibhodi yabaphambili ngokuvulekile: Ikhadi le-MOSS-Transcribe-preview-2B lithi imodeli ilungiswe kahle ngokufunda okuqiniswayo ekuhlukaniseni ukuqeqeshwa kwebhodi yabaphambili ye-Open ASR. Lokho kuyadalulwa, okungaphezu kokuningi, kodwa kusho ukuthi amaphuzu akala ibhentshimakhi kunekhono.

Idatha yethrekhi yangasese ihlela kabusha ibhodi: I-Appen inikele ngamasethi okuhlola abambezelekile afaka ama-accents ase-Australia, e-Canada, e-Indian, nawaseMelika ezimeni ezibhaliwe nezingxoxo. Uma lawo masethi ayimfihlo eshintshwa avuliwe, i-zoom/scribe_v1 isuka ku-#4 iye ku-#1 bese umholi webhodi labaphambili lomphakathi ehlisa isikhundla. Amamodeli alungiselwe inkulumo efundeke kahle ehlisa isithunzi ngokuzenzakalelayo kumsindo wengxoxo.

Sebenzisa ibhodi yabaphambili ukuze wakhe uhlu olufushane. Ungayisebenzisi ukukhetha owinile.

Isigaba sokunemba

I-Cohere Transcribe (I-2B, i-Apache 2.0, izilimi ezingu-14) iyimodeli empeleni ethunyelwe ekukhiqizweni. Ilandwe izikhathi ezingaphezu kuka-620,000 enyangeni edlule futhi inosekelo lwesikhathi sokusebenza kuyo yonke indawo transformersvLLM, mlx-audio ye-Apple Silicon, i-Rust port, kanye nokwakhiwa kweWebGPU. Kuyi-Conformer encoder ene-Transformer decoder engasindi, eqeqeshwe kusukela ekuqaleni. U-Cohere uphinde waqhuba nokuhlola okuthandwayo komuntu, lapho izichasiselo eziqeqeshiwe zathola khona okulotshiweyo ukuze kugcinwe incazelo, ukubona izinto ezingekho, kanye nezinhlangano eziqanjwe amagama: isilinganiso esingu-61% esimaphakathi sokuwina, 78% siqhathaniswa ne-IBM Granite 4.0 1B Speech kanye no-64% ngokumelene ne-Whisper enkulu-v3.

Ingxenye yemikhawulo yekhadi layo eliyimodeli ithembekile ngokungajwayelekile futhi kufanele ifundwe ngaphambi kokuzibophezela. Akukho ukutholwa kolimi okuzenzakalelayo, azikho izitembu zesikhathi, futhi akukho mshini wokudayela, futhi imodeli izimisele ukuloba ukuthula, ngakho u-Cohere uncoma ukuthi kulungiselelwe i-VAD noma isango lomsindo. I-repo iphinde ibe ngemuva kwesivumelwano solwazi lokuxhumana naphezu kwelayisensi ye-Apache 2.0.

Inkulumo Yegwadle 4.1 2B (2B, Apache 2.0) ukukhetha okungcono kakhulu uma udinga amandla kunenombolo ephansi. Izilimi eziyisithupha ze-ASR kanye nokuhunyushwa kwenkulumo eqondiswe kabili, ukuchema kohlu lwegama elingukhiye lwamagama nejagoni, izimpawu zokubhala kanye nombhalo oyiqiniso ohlanganisa osonhlamvukazi bebizo lesiJalimane. Uqeqeshwe ngamahora angu-174,000. I-RTFx 231.29. I-IBM iphinde ithumele izingane zakubo ezimbili: -plus yengeza i-ASR ehlobene nesipika kanye nezitembu zesikhathi zezinga legama, kanye -nar kuxoxwa ngezansi.

I-Canary-Qwen-2.5B (2.5B, CC-BY-4.0, IsiNgisi) ibhanqa isishumeki se-FastConformer nesikhiphi se-Qwen3-1.7B futhi sisebenza ngezindlela ezimbili — ukuloba okumsulwa, noma imodi ye-LLM lapho idikhoda ifinyeza futhi iphendule imibuzo mayelana nokulotshiwe. 5.63% WER ku-RTFx 418. Qaphela ukuthi i-AMI yeqiwe yafinyelela cishe ku-15% wedatha yokuqeqeshwa, echemile okukhiphayo ekulobeni okulotshiweyo okulondolozwa kokungaguquguquki kwezwi nezwi. Leso isici somsebenzi wezomthetho kanye nenkathazo kumanothi omhlangano.

I-Qwen3-ASR-1.7B (I-Apache 2.0) ihlanganisa izilimi nezilimi zesigodi ezingu-52 — izilimi ezingu-30 kanye nezilimi zesigodi zesiShayina ezingu-22 — ngo-5.76%. Ithunyelwa nekhithi yamathuluzi egcwele yokukhomba kanye nemodeli ehlukile yokuqondisa ngempoqo yezitembu zesikhathi ngezilimi eziyi-11. Kunoma yini ethinta isiMandarin noma inkulumo yesifunda yesiShayina lesi isiqalo esisobala.

Isigaba sokuphuma

Ukunemba phezulu kwenkundla manje kuyahlukahluka ngephoyinti elilodwa le-WER. Ukukhipha imali kuyahlukahluka ngaphezu kokuhleleka kobukhulu, okusho ukuthi okuphumayo kuvame ukunquma i-invoyisi.

I-Parakeet TDT 0.6B v3 (0.6B, CC-BY-4.0) ingumholi we-throughput phakathi kwamamodeli avuliwe ngezilimi eziningi ku-RTFx 3332.74 kuzo zonke izilimi zase-Europe ezingama-25 ezinomazisi wolimi ozenzakalelayo, ophatha imizuzu engama-24 ngokudlula okukodwa ku-A100 80GB. Ibiza u-6.32% WER — cishe iphuzu elilodwa ngaphezu kweGranite 4.1 2B cishe izikhathi eziyishumi nane zomsindo nge-GPU-yesibili ngayinye.

Inkulumo Yegwadle 4.1 2B-NAR umphumela wobunjiniyela onentshisekelo kukho. Ayina-autoregressive: ihlela i-hypothesis ye-CTC ekudluleni okukodwa kokuya phambili isebenzisa i-LLM eqondiswa kabili, ifinyelela ku-RTFx ~1820 ku-H100 eyodwa kusayizi weqoqo 128. Iyeka isiJapane, ukuhumusha inkulumo, nokuchema kwegama elingukhiye ukuze ufike lapho.

Qwen3-ASR-0.6B igcina zonke izilimi ezingama-52 futhi ifinyelela ku-2000× throughput at concurrency 128.

Isigaba sokusakaza

Inqwaba ye-WER ibonakala iyisivivinyo esingalungile semodeli yokusakaza-bukhoma, futhi ibhodi yabaphambili iyawathola noma kunjalo. I-Voxtral Realtime ihlezi kokuthi 7.68% kanye ne-Kyutai STT 2.6B kokuthi 6.40% — kokubili ngaphansi kwe-Whisper — futhi ayikho inombolo ekutshela okuthile okuwusizo mayelana nokusetshenziswa kwayo okuhlosiwe.

I-Voxtral Mini 4B Realtime 2602 (I-Apache 2.0, izilimi ezingu-13) imodeli yolimi engu-3.4B kanye nesishumeki somsindo esiyisizathu esingu-970M esiqeqeshwe kusukela ekuqaleni, nokunaka kwewindi elislayidayo kuwo womabili uhhafu ukuze usakaze ngempumelelo ngaphandle komkhawulo. Ukubambezeleka kokulotshiweyo kuyalungiseka ngezinyathelo ezingu-80ms ukusuka ku-80ms ukuya ku-1200ms, kanye nenketho ezimele engu-2400ms; I-Mistral incoma i-480ms njengendawo emnandi futhi ibika ukuthi kuleso silungiselelo ifana namamodeli avulekile angaxhunyiwe ku-inthanethi. Isebenza nge-16GB GPU eyodwa futhi inokusekelwa kwe-day-0 vLLM Realtime API.

I-Kyutai STT (CC-BY-4.0) iza ngezimo ezimbili: imodeli engu-~1B yesiNgisi/yesiFulentshi enokubambezeleka okungu-0.5s kanye nesitholi somsebenzi wezwi le-semantic esakhelwe ngaphakathi, kanye nemodeli yesiNgisi engu-2.6B kuphela enokulibaziseka okungu-2.5s. Kwabenzeli bezwi, i-VAD ye-semantic ibaluleke ngaphezu kokulibaziseka kokuloba – ibikezela ukuthi isipikha sesiqede nini ngempela, okuyikhona obusa ukubambezeleka okubonwayo. I-H100 inikeza ukusakaza okuhambisanayo okungu-400 ngesikhathi sangempela.

Isigaba sokumboza

I-Meta's Omnilingual ASR (I-Apache 2.0, corpus CC-BY) ayiqhudelani nge-WER futhi akufanele ihlolwe njengokungathi ihlolwa. Ihlanganisa izilimi ezingu-1,600+ ngokomdabu futhi inwebela ku-5,400+ ngokufunda ku-zero-shot-in-context, eyakhelwe kusishumeki se-wav2vec 2.0 esikalelwe ku-7B futhi saqeqeshwa kusengaphambili cishe amahora angu-4.3M. Okuhlukile kwe-7B LLM-ASR kufinyelela izinga lephutha lezinhlamvu elingaphansi kuka-10% ku-78% yezilimi ezisekelwayo, kuhlanganisa nezingu-500+ ezingakaze zisetshenziswe ngaphambilini yinoma iyiphi isistimu ye-ASR. Osayizi besifaki khodi baqala ku-300M ukuya ku-7B. I-Meta iphinde yakhipha i-Omnilingual ASR Corpus ehlanganisa izilimi ezingafinyeleleki ezingu-350+.

Ukuhleba okukhulu-v3 (1.55B, MIT, izilimi ezingu-99) ifinyelelwe ngokunemba cishe ngamamodeli avuliwe ayishumi, futhi ihlala iyiphutha elilungile lesigaba esikhulu samaphrojekthi. I-MIT ilayisense efakwe kancane kancane kulo mkhakha. I-runtime ecosystem — whisper.cpp, faster-whisper, WhisperX — ayinakho okulinganayo phakathi kokukhishwa okusha. Uma imfuneko yakho “iwulimi oluthile, ihadiwe elithile, akekho ummeli welayisensi,” kuseyimpendulo.

Unqenqema locwaningo

i-diffusion-gemma-asr-encane kusuka ku-YC startup I-Interfaze iwukukhululwa kobuciko bokwakha obuningi bonyaka. Ikhiqiza okulotshiweyo ngokusakazwa okufanayo okuchaza ngekhanvasi enamathokheni angama-256 ezinyathelweni eziyi-8 kuye kweziyi-16, ngakho-ke izindleko zokuqopha azikhuli ngobude bombhalo. Kuqeqeshwe amapharamitha angu-~42M kuphela – 0.16% wezisindo – phezu kwe-26B DiffusionGemma efriziwe kanye nesishumeki sekhodi esifriziwe. Ifinyelela ku-6.6% WER ekuhlolweni kwe-LibriSpeech cishe ku-11 kuya ku-17× ngesikhathi sangempela. Futhi ifinyelela ku-15.7% ku-FLEURS English kanye no-29.6% CER ku-FLEURS Mandarin, ngakho phatha isibalo se-LibriSpeech njengosilingi. Akekho umuntu okufanele asebenzise lokhu. Wonke umuntu osebenza ku-ASR kufanele ayifunde.

I-MOSS-Transcribe-Diarize 0.9B (I-Apache 2.0, izilimi ezingu-50+) ixazulula inkinga abaningi abayizibayo: ikhipha amalebula esipikha, izitembu zesikhathi samagama, nokulotshiwe esizukulwaneni esisodwa esikhundleni sokuhlanganisa i-ASR kusitaki sokudaya esihlukile. 128k umongo, cishe imizuzu engu-90 yomsindo ngephasi eyodwa, i-RTF ~0.017 ku-RTX 4090, nokuchema kwe-hotword.

Ilayisensi iyahlukaniswa akekho umuntu oyifundayo kuze kube iyasetshenziswa

Lesi yisigaba esivimba ukuthunyelwa, futhi inkambu ihlukanisa ngokuhlanzekile:

I-Apache 2.0: Cohere Transcribe, Granite Speech 4.1 (zonke izinhlobo ezintathu), Qwen3-ASR (bobabili osayizi), Voxtral Mini Realtime, Omnilingual ASR, ARK-ASR, MOSS-Transcribe. Asikho isibopho semfanelo, ukusetshenziswa kwezohwebo akukhawulelwe. Qaphela ukuthi i-repo ka-Cohere ifakwe ngemuva kwesivumelwano solwazi lokuxhumana nakuba ilayisense ngokwayo kuyi-Apache 2.0.

I-MIT: Hleba kakhulu-v3. Inketho evumela kakhulu ensimini.

CC-BY-4.0: Canary-Qwen-2.5B, Parakeet TDT 0.6B v3, Kyutai STT. Kuyasebenziseka ngokwentengiso, kodwa isichasiso siyadingeka. Ngomkhiqizo oshumekiwe noma i-API enelebula elimhlophe, lokho kuyisibopho sokuthobela sangempela, futhi isizathu esisodwa esivame kakhulu ukuthi amaqembu agcine ethumele imodeli okungeyona enembe kakhulu ayihlolile.

I-Meta's Omnilingual ASR ihlukanisa lezi ezimbili: I-Apache 2.0 yamamodeli, i-CC-BY yekhophasi.

Indlela yokukhetha ngempela

Qalisa leli oda, hhayi nje i-oda lebhodi labaphambili:

1. Ilayisensi: Uma isichasiso yisivimbeli, i-CC-BY-4.0 isusa i-Parakeet, i-Canary-Qwen, ne-Kyutai ngaphambi kokuthi ulinganise noma yini.
2. Ukusabalala kolimi: Izilimi ze-Cohere's 14, Granite's 6, kanye ne-Canary's English-kuphela imikhawulo eqinile, hhayi ethambile. I-Cohere ngaphezu kwalokho ayinakho ukubona ngokuzenzakalelayo ulimi, ngakho-ke kufanele wazi ulimi kusenesikhathi.
3. Ukusakaza noma iqoqo: Lokhu okokwakha. Alikho inani lokushuna elishintsha isikhiphi sekhodi esingaxhunyiwe ku-inthanethi sibe imodeli yokusakaza yokubambezeleka ephansi.
4. Bese ukala i-WER kumsindo wakho: Ukusabalala phakathi kwamamodeli ayishumi aphezulu kubhentshimakhi yomphakathi kungaphansi kwephuzu elilodwa. Ukusabalala kumsindo wakho onesigcino, onomsindo, nesizinda esithile kuzoba izikhathi ezimbalwa lokho, futhi ngeke kuklelise amamodeli ngohlelo olufanayo.
5. Bese ubala izindleko ngehora ngalinye lomsindo kuma-GPU akho: Izibalo ze-RTFx zikalwa ngosayizi abakhulu beqoqo ku-hardware yedathacenter futhi azidluliseli.