InoteveraGaidhi rinotevera
Batch vs Real-Time Inference
Tekinoroji
Nhungamiro yehunyanzvi
Pitch inoshanduka shanduko yakaonekwa yakakosha frequency uku uchivavarira kuchengetedza nguva, nepo nguva inotambanudzwa inochinja nguva uchivavarira kuchengetedza kukwirira.
Teerera kugwaro iriAI-yakarondedzerwa · ~3 min
Phase vocoder, pitch-synchronous nzira dzakadai sePSOLA, uye neural masisitimu anosvika pamabasa aya zvakasiyana, ane zvigadzirwa uye kukodzera zvinoenderana nechiratidzo.
Kuchinja kwepitch uye kutambanudza nguva kune hukama asi shanduko dzakasiyana. Shanduko yakapfava muyero yekutamba inoshandura zvese nguva uye kukwirira: kutamba nekukasira kunopfupisa clip uye kusimudza mutsindo wayo, uku kutamba kunononoka kuchirebesa nekudzikisa kukwirira. Muchinjiziri wenzwi anoedza kushandura nzwi pasina kushandura nguva; chitanda chenguva chinoedza kushandura nguva pasina kushandura inzwi. Zvinangwa izvi zvinoda kugadziridzwa kwemasaini kupfuura kushandura sampuli-reti label. A phase vocoder inoshanda muFourier transform domain yenguva pfupi. Inoshandura kupatsanurwa kwemafuremu ekuongorora kuti ichinje nguva, yobva yagadzirisa kufambira mberi kwechikamu kuti igadzire inowirirana synthesis. Inogona kubata zvakajairika odhiyo asi inogona kuzora zvinopfuudza kana kugadzira phasiness, kunyanya nezvinhu zvinorova kana zvakanyanya kutambanudza zvinhu. Kuchinja kwepitch kunogona kuvakwa nekubatanidza chiyero chenguva-chikero nekuenzanisazve, asi ruzivo rwunoona zvigadzirwa uye muripo wenguva. Pitch-synchronous overlap-add nzira dzinobatanidza kugadzirisa kune inofungidzirwa nguva yekudzika. PSOLA inogona kushanda nemazvo pakutaura kwekutaura kana mapitch mamaki achivimbika, zvichiita kuti zvive zvakakosha pakuchinja kurira kana nguva uchichengetedza kutaura-kunge waveform. Haikodzeri zvakananga kune ruzha rusina kunzwikwa kana mimhanzi yakasarudzika. Neural nzira dzinodzidza shanduko kubva kudhata uye dzinogona kuburitsa mhedzisiro yezvinyorwa zvinotsigirwa, asi zvinoda mamodheru akadzidziswa uye anogona kushandura ruzivo kana kukundikana kunze kwemamiriro avo ekudzidziswa. Hapana nzira ine artifact-yemahara kune yega chiratidzo. Transients, vibrato, polyphony, kutaura kune ruzha, uye hombe shanduko zvinhu zvinofumura kusasimba kwakasiyana. Iyo algorithm yakagadziridzwa yezwi haigone kuchengetedza chigubhu chinorova kana mimhanzi timbre. Ongorora nekuteerera uye chinangwa chebasa matanho anoratidza nyaya yekushandisa. Chengetedza data rakashandurwa mukati mezviyero zvinonzwisisika kana uchiwedzera mienzaniso yekudzidziswa kuitira kuti mabhii arambe achishanda. Maraibhurari anofumura mashandiro ane akasiyana parameter magungano. Taura kuti chiyero chinomiririra nguva yekubuda inoenderana nekuisa, uye ona kana basa racho richichengetedza nzwi nekumisikidza. Rekodha iyo algorithm, vhezheni, sampling rate, uye marongero. Gara uchitarisa nguva yekubuda, kutema, uye muganho wekubata usati wasanganisa yakashandurwa odhiyo kuita dhatabheti kana kugadzirwa kwekufambisa.
Vhidhiyo mupepeti inononoka kurondedzera uku ichichengetedza inzwi rayo kuitira kuti mutauri arege kunzwika akaderera zvisingaite.
Chishandiso chemumhanzi chinoburitsa rwiyo rupfupi uchichengeta nguva yarwo, chozotarisa kuti pane smearing yenguva pfupi uye shanduko yetimbral.
Matauriro synthesis sisitimu inoshandisa pitch-synchronous overlap-add pakutaura kwezwi uye inobata matunhu asina kutaurwa zvakasiyana.
Muongorori wekuteerera anofananidza chikamu-vocoder yakatambanudzwa neneural nzira yekurova uye toni dzakasimba.
Maturusi ekushandura maturusi anoenderera mberi achibatanidza nzira dzekugadzirisa-signal neural chizvarwa kudzikisa zvigadzirwa paizwi, mimhanzi, uye ruzha rwezvakatipoteredza. Nzira dzakadzidzwa dzinogona kubata zvimwe zvinhu zvinonetsa zvirinani, nepo classical phase vocoder inoramba ichishanda uye ichidzoreka. Vashandisi vanofanirwa kutarisira kuti mhando ichasiyana nezvirimo uye huwandu hweshanduko. Ongororo dzinogona kubatsirikana kubva mukuteerera bvunzo pamwe chete nenguva, kukwirira, uye kudzika-basa matanho pane kuvimba nechibodzwa-chinangwa chimwe chete. Miedzo inofanirwa kuvhara toni dzakasimba uye dzinopfuura dzakapinza. Matanidza shanduko yezvinhu kune yekushandisa kesi.
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
Pitch inoshanduka shanduko yakaonekwa yakakosha frequency uku uchivavarira kuchengetedza nguva, nepo nguva inotambanudzwa inochinja nguva uchivavarira kuchengetedza kukwirira. Phase vocoder, pitch-synchronous nzira dzakadai sePSOLA, uye neural masisitimu anosvika pamabasa aya zvakasiyana, ane zvigadzirwa uye kukodzera zvinoenderana nechiratidzo.
Kuchinja kwepitch kune chinangwa chekushandura nzwi pasina kushandura kureba kweklip.
Nguva inotambanudzwa inoshandura nguva uchiyedza kuchengetedza izwi rekutanga.
Kutamba nekukurumidza kana kunonoka kunoshandura kumhanya kwenguva uye kunoonekwa frequency pamwechete.
Phase vocoder inoshandura nguva yenguva uye chikamu mune STFT inomiririra.
Kurwiswa kwenguva pfupi kunogona kusvibiswa nefuremu-based phase-vocoder processing.
Ramba uchidzidza
Mamwe madhairekitori akasarudzirwa nyaya iyi
InoteveraGaidhi rinotevera
Batch vs Real-Time Inference
Tekinoroji