Mozilla Inotangisa Yakadzika Speech 0.9 Speech Kuziva Injini

YakadzikaSpeech1

Kutanga kwakaburitswa Injini yekuziva izwi DeepSpeech 0.9 yakagadzirwa naMozilla, iyo inoshandisa kuvakwa kwe kuzivikanwa kwezwi wezita rimwe chete rakatsanangurwa neBaidu vaongorori.

Kuitwa yakanyorwa muPython uchishandisa muchina wekudzidza chikuva TensorFlow uye inogoverwa pasi pemahara MPL 2.0 rezinesi.

Nezve DeepSpeech

DeepSpeech ine maviri masystems: yeacoustic modhi uye dhodhi. Iyo acoustic modhi inoshandisa yakadzika michina yekudzidzira matekiniki kuverenga mukana wekuti vamwe vatambi varipo mune yekuisa ruzha.

Iyo decoder inoshandisa ray yekutsvaga algorithm kushandura iyo hunhu hunogona dhata kuva mameseji anomiririrwa. DeepSpeech iri nyore kupfuura echinyakare masisitimu uye panguva imwechete inopa yakakwira mhando yekuzivikanwa pamberi pekunze ruzha.

Iko kusimudzira hakushandise echinyakare acoustic mamodheru uye pfungwa yemaphonimesi; pachinzvimbo, yakanyatsogadziridzwa neural network-yakavakirwa muchina wekufunda muchina inoshandiswa, iyo inobvisa kudiwa kwekukudziridza zvikamu zvakasiyana kuti zvienzanise zvakashata zvakasiyana senge ruzha, maungira, uye maitiro ekutaura.

Iyo kit inopa akadzidziswa mamodheru, sampuro inonzwika mafaira uye raira matambo ekuziva maturusi.

Iyo yakapera modhi inopihwa yeChirungu neChinese chete. Kune mimwe mitauro, unogona kudzidza iyo system pachako maererano nemirairo yakanamatira, uchishandisa iro datha rezwi rakaunganidzwa neyakajairika Izwi purojekiti.

When iyo yakagadzirira-kushandisa-modhi yemutauro weChirungu unopihwa kurodha pasi unoshandiswa, huwandu hwezvikanganiso zvekuziva muDepSpSpoti ndeye 7.06% painotariswa uchishandisa iyo LibriSpeech bvunzo suite.

Kuti uenzanise, chiyero chekukanganisa kwemunhu chinofungidzirwa pa5,83%.

Mumuenzaniso wakarongedzerwa, mhedzisiro yakanakisa yekuzadzikiswa inowanikwa nekurekodha kwakachena kwezwi rechirume neizwi reAmerica munzvimbo isina ruzha rwekunze.

Sekureva kwemunyori weVosk Continuous Speech Recognition Raibhurari, zvinokanganisa iyo Izwi Rinoiswa zvakasarudzika ndeyekurerekera kudivi rekutaura (hukuru hwevarume vane makore makumi maviri nemakumi matatu uye kushomeka kwezvinhu nezwi revakadzi, vana nevakuru), kushomeka kweshanduko yemazwi (kudzokororwa kwemitsara imwe chete) uye kugoverwa kwe MP20 marekodhi anokanganisa kukanganisa.

Kukanganisa kweDeepSpeech kunosanganisira kusashanda zvakanaka uye yakanyanya ndangariro yekudyiwa muodhiyo, pamwe nezvakakosha zviwanikwa kudzidzisa modhi (Mozilla inoshandisa sisitimu ine 8 Quadro RTX 6000 GPU ine 24GB VRAM mune yega yega).

Iyo yekudzikira kune iyi nzira ndeyekuti yekuremekedzwa kwemhando yepamusoro uye kudzidziswa kweiyo neural network, Injini yeSpainSpeech inoda yakawanda data kusanzwisisika kwakarairwa mumamiriro chaiwo nemanzwi akasiyana uye pamberi pezvaiitika ruzha.

Iyi data yakarongedzwa neCommunity Voice chirongwa chakagadzirwa muMozilla, iyo inopa yakasimbiswa data yakagadzirirwa nemaawa 1469 muChirungu, 692 muchiGerman, 554 muchiFrench, maawa 105 muchiRussia uye maawa makumi maviri nemaviri muUkraine.

Kana uchidzidzisa yekupedzisira yeChirungu modhi yeDeepSpeech, kuwedzera kune Yakajairika Izwi, data kubva kuLibriSpeech, Fisher uye switchchboard mapurojekiti anowedzerwa anoshandiswa, pamwe neanenge maawa mazana gumi nemanomwe ekurekodha ezvirongwa zveredhiyo.

Pakati pekuchinja kubazi idzva, mukana wekumanikidza huremu hwemashoko hunoratidzwa yakasarudzwa panguva yekugadzirisa maitiro.

Inotaridzawo rutsigiro rweiyo Electron 9.2 chikuva uye sarudzo yekumisikidza yeiyo dura yekumisikidza mashandiro (Layer Norm) kana uchidzidzisa neural network

Dhawunirodha uye tora

Kuita kwacho kwakaringana kushandisa mota muLePotato, Raspberry Pi 3 uye Raspberry Pi 4 mabhodhi, pamwe neGoogle Pixel 2, Sony Xperia Z Premium uye Nokia 1.3 mafoni.

Akagadzirira ma module anopihwa kushandisa Python, NodeJS, C ++, uye .NET kubatanidza kutaura kunozivikanwa kuita muzvirongwa zvako (vechitatu-bato vagadziri vakazvigadzirira zvakasiyana maRust, Go, uye V).


Iva wekutanga kutaura

Siya yako yekutaura

Your kero e havazobvumirwi ichibudiswa. Raida minda anozivikanwa ne *

*

*

  1. Inotarisira iyo data: Miguel Ángel Gatón
  2. Chinangwa cheiyo data: Kudzora SPAM, manejimendi manejimendi.
  3. Legitimation: Kubvuma kwako
  4. Kutaurirana kwedata
  5. Dhata yekuchengetedza: Dhatabhesi inobatwa neOccentus Networks (EU)
  6. Kodzero: Panguva ipi neipi iwe unogona kudzora, kupora uye kudzima ruzivo rwako