InoteveraGaidhi rinotevera
Voicebox Flow-Matching Speech Generation
AI Yenyaya
Chidzidzo cheAI yeRuzha
WaveGlow inoyerera-yakavakirwa neural vocoder kubva kuNVIDIA iyo inogadzirisa mafungu ekutaura kubva kumel-spectrograms mune imwechete pass pasina autoregression.
Teerera kugwaro iriAI-yakarondedzerwa · ~2 min
It matters because it delivers high-quality audio faster than real time using only a simple likelihood loss.
WaveGlow, yakaburitswa naPrenger, Valle, uye Catanzaro kuNVIDIA muna 2018, inosanganisa mazano kubva kuGlow neWaveNet kuvaka vocoder inokurumidza uye iri nyore kudzidzisa. Kusiyana nemavokodha eGAN, inoyerera yakajairwa: inodzidza mepu isingachinjiki pakati pekugovera kwakapusa kweGaussian uye odhiyo waveform, yakarongedzwa pane mel-spectrogram. Kudzidzira kunowedzera iyo chaiyo log-mukana weiyo data, saka haidi akaparadzana akaparadzana, hapana auto-regression, uye hapana maviri-network mudzidzisi-mudzidzi distillation iyo yapfuura yakafanana WaveNet nzira dzinodiwa. Kuti ugadzire odhiyo iwe unotevedzera Gaussian ruzha uye mhanyisa iyo invertible network kumashure. WaveGlow inogadzira kutaura kwemhando inofananidzwa neWaveNet uku ichigadzira nekukurumidza kupfuura nguva chaiyo paGPU yemazuva ano.
Kubatanidza neTacotron 2 muNVIDIA's reference TTS pombi yekugadzira yakasikwa studio-yemhando yekutaura
Yekukurumidza GPU yekutaura synthesis yekurondedzera, dubbing, uye zvemukati kugadzira workflows
Kugadzira kudzidziswa uye demo odhiyo mukutsvagisa uko kwakagadzikana, kumwe-kurasikirwa kudzidziswa kunosarudzwa
Chaiyo-nguva-inokwanisa izwi kubuda mumashandisirwo masisitimu anomhanya paNVIDIA hardware
WaveGlow yakaratidza kuti yakachena kuyerera mavhokodha anogona kukwikwidza autoregressive mhando, ichipesvedzera gare gare kuyerera uye kuyerera-kunoenderana odhiyo modhi. Kureruka kwaro-kurasikirwa kunoramba kuchifadza, kunyange GAN vocoders seHiFi-GAN ikozvino inowanzokunda pakukura nekumhanya. Kutarisa kumberi, kuyerera-kwakavakirwa uye kuyerera-kufananidza mazano ari kusimuka mune yemazuva ano kupararira-yakatarisana neTTS, uye WaveGlow-maitiro invertible madhizaini anoenderera mberi nekuzivisa tsvakiridzo pane chaiyo-inogoneka, inodzoreka, uye inoshanda waveform chizvarwa.
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
WaveGlow inoyerera-yakavakirwa neural vocoder kubva kuNVIDIA iyo inogadzirisa mafungu ekutaura kubva kumel-spectrograms mune imwechete pass pasina autoregression. Izvo zvine basa nekuti inoburitsa emhando yepamusoro odhiyo nekukurumidza kupfuura nguva chaiyo uchishandisa chete nyore kurasikirwa.
WaveGlow inosanganisa iyo invertible kuyerera chimiro cheGlow ine odhiyo-modhiyo dhizaini yeWaveNet.
WaveGlow ndeye yakajairika kuyerera iyo inodzidza invertible mepu pakati peGaussian ruzha uye odhiyo.
Nekuti iyo network haichinjike, unodhonza Gaussian ruzha uye unomhanyisa kuyerera uchidzokera kumashure, wakamisikidzwa pane mel-spectrogram.
Sekuyerera, WaveGlow inowedzera iyo chaiyo log-inogona, inoda kusarura kana distillation.
Yese inoyerera nhanho inobatanidza affine coupling layer ine invertible 1x1 convolution yakatorwa kubva kuGlow.
Ramba uchidzidza
Mamwe madhairekitori akasarudzirwa nyaya iyi
InoteveraGaidhi rinotevera
Voicebox Flow-Matching Speech Generation
AI Yenyaya