SoundStream Neural Codec
SoundStream bụ Google's end-to-end neural audio codec nke na-akpakọ okwu na egwu na bitrate dị oke ala ma na-echekwa ịdịmma.
Nchịkọta
SoundStream is Google's end-to-end neural audio codec that compresses speech and music to extremely low bitrates while preserving quality. Ọ dị mkpa n'ihi na ọ na-akụ codecs ọdịnala dịka Opus n'otu bitrate ma na-akwado ụdị ọdịyo ọgbara ọhụrụ.
SoundStream Neural Codec na-anọdụ n'usoro ọdịyo-AI nke na-agbanwe okwu, egwu, na ụda maka nzikọrịta ozi, nnweta na mmepụta mgbasa ozi.
Ime miri emi
Nke __AIU_PROTECTED_11_ webatara na 2021, SoundStream bụ codec neural zuru oke nke e wuru site na iberibe atọ a zụrụ ọnụ: ihe ngbanwe mgbanwe nke na-atụgharị raw waveform ka ọ bụrụ usoro kọmpat nke vectors, residual vector quantizer (RVQ) nke na-agbagha vectors ndị ahụ, yana ihe na-agbanwe agbanwe. A zụrụ ya site na mfu nrụzigharị yana onye ịkpa oke n'ụdị GAN, yabụ mmepụta na-ada ụda eke kama ịbụ nso nsoro ọnụ. Ngosipụta pụtara bụ 'scalable' ma ọ bụ ọzụzụ quantizer-dropout: otu ụdị nwere ike ịrụ ọrụ n'ofe bitrates site na ihe dị ka 3 ruo 18 kbps naanị site na iji ọkwa quantizer karịa ma ọ bụ dị nta na ntinye, na-enweghị nlọghachi azụ. Na 3 kbps, a na-akọ na ọ karịrị Opus na 12 kbps n'ime ule ige ntị, ijikwa okwu, egwu, na ọdịyo izugbe n'otu ụdị nke nwere ike ịgba ọsọ ozugbo na CPU smartphone.
Nghọta nka nka
Ụdị ebili mmiri ahụ na-agafe mgbanwe ndị gbawara agbawa na-agbadata nke ukwuu, na-ewepụta otu ntinye n'otu etiti (dịka freelụ 75/sekọnd). RVQ wee tinye koodu ntinye ọ bụla dị ka njupụta nke indices codebook. Bitrate ha nhata nha nha n'etiti ọnụọgụ nke oge quantizer na-arụ ọrụ n'otu akwụkwọ koodu. Quantizer dropout randomly truncates the RVQ stack during training, forcing earlier codebooks to carry the most important information so the codec degrades gracefully at lower rates.
Ịmalite Codec Neural SoundStream
Iji wuo nghọta miri emi, na-emeso SoundStream Neural Codec dị ka ihe nlereanya na-arụ ọrụ, ọ bụghị otu njirimara. Kọwaa nsonaazụ achọrọ, kọwapụta echiche, ma kewaa ihe sistemụ nwere ike ime nke ọma na ihe ka na-achọ mkpebi ndị ọkachamara.
In practice, strong teams using SoundStream Neural Codec treat quality, latency, and consent as equally important parts of the deployment strategy. Ha na-edepụta njirisi ịga nke ọma nke ọma, nwalee megide data ziri ezi yana usoro ọrụ, yana na-atụgharị dabere na usoro ọdịda ahụrụ karịa karịa mmeri otu oge. Nke a bụ ebe nghọta usoro ihe atụ na-atụgharị ghọọ ike na-adịgide adịgide n'ofe ngwaahịa, amụma na arụmọrụ.
Ọ na-eme ka nnweta ya dịkwuo mma site na ndegharị, ịkọ akụkọ, na ntụgharị olu. N'otu oge ahụ, iji olu eme ihe n'ụzọ na-ezighị ezi na ihe egwu mpụta ga-abawanye mgbe nkwenye na-efu. Ụzọ kachasị na-agbanwe agbanwe bụ ijikọ ọsọ nnwale na ịdọ aka ná ntị ọchịchị: ndị na-anya ụgbọ elu, ijide ihe akaebe, bipụta ndekọ mkpebi, na na-aga n'ihu na-emelite nchekwa dị ka omume nlereanya, atụmanya ndị ọrụ, na ihe iwu chọrọ.
Mmetụta atụmatụ
Ọ na-eme ka nnweta ya dịkwuo mma site na ndegharị, ịkọ akụkọ, na ntụgharị olu.
Ọ na-eme ka nnweta ya dịkwuo mma site na ndegharị, ịkọ akụkọ, na ntụgharị olu. N'ịkwanye ọkwa dị elu, a na-atụgharị nke a ka ọ bụrụ iwu arụ ọrụ enwere ike ịtụnye, oke nwe, na emume ntụlegharị ugboro ugboro ka ndị otu wee nwee ike ịbawanye ntụkwasị obi kama iwelite enweghị mgbagha.
Ndị otu mgbasa ozi nwere ike ibubata ọdịyo a na-egbu maramara ngwa ngwa site na iji obere mmefu ego.
Ndị otu mgbasa ozi nwere ike ibubata ọdịyo a na-egbu maramara ngwa ngwa site na iji obere mmefu ego. N'ịkwanye ọkwa dị elu, a na-atụgharị nke a ka ọ bụrụ iwu arụ ọrụ enwere ike ịtụnye, oke nwe, na emume ntụlegharị ugboro ugboro ka ndị otu wee nwee ike ịbawanye ntụkwasị obi kama iwelite enweghị mgbagha.
Sistemụ na-eche ihu ndị ahịa nwere ike hazie mkparịta ụka n'ọtụtụ buru ibu.
Sistemụ na-eche ihu ndị ahịa nwere ike hazie mkparịta ụka n'ọtụtụ buru ibu. N'ịkwanye ọkwa dị elu, a na-atụgharị nke a ka ọ bụrụ iwu arụ ọrụ enwere ike ịtụnye, oke nwe, na emume ntụlegharị ugboro ugboro ka ndị otu wee nwee ike ịbawanye ntụkwasị obi kama iwelite enweghị mgbagha.
Mmejuputa n'ezie n'ụwa
Na-akpakọ oku olu na ~ 3 kbps ka ọ na-ada nke ọma karịa codecs nketa na bitrates dị elu
Na-emepụta ihe ngosi ọdịyo pụrụ iche na-enye Google's AudioLM na ụdị mmepụta MusicLM
Mgbasa ụda ọdịyo dị obere obere oge na ngwaọrụ mkpanaka nwere ntinye na ngbanwe on-CPU
Na-echekwa ma ọ bụ na-ebusa egwu na ụda ihe nke ọma n'otu ụdị nke na-ejikwa ụdị ọdịnaya niile
Usoro mmejuputa
SoundStream Neural Codec na omume
Na-akpakọ oku olu na ~ 3 kbps ka ọ na-ada nke ọma karịa codecs nketa na bitrates dị elu.
Otu dị iche iche na-enwetakarị nsonaazụ ka mma mgbe ha na-akọwapụta ọnụ ụzọ dị mma n'ihu, na-eme ka ụzọ mmadụ si abawanye maka oke ikpe, ma soro ma uru nrụpụta yana ụgwọ njehie n'ime oge.
SoundStream Neural Codec na omume
Na-emepụta ihe ngosi ọdịyo pụrụ iche nke na-enye Google's AudioLM na ụdị mmepụta MusicLM.
Otu dị iche iche na-enwetakarị nsonaazụ ka mma mgbe ha na-akọwapụta ọnụ ụzọ dị mma n'ihu, na-eme ka ụzọ mmadụ si abawanye maka oke ikpe, ma soro ma uru nrụpụta yana ụgwọ njehie n'ime oge.
SoundStream Neural Codec na omume
Mgbasa ụda ọdịyo dị obere obere oge na ngwaọrụ mkpanaka nwere ntinye na ngbanwe na CPU.
Otu dị iche iche na-enwetakarị nsonaazụ ka mma mgbe ha na-akọwapụta ọnụ ụzọ dị mma n'ihu, na-eme ka ụzọ mmadụ si abawanye maka oke ikpe, ma soro ma uru nrụpụta yana ụgwọ njehie n'ime oge.
SoundStream Neural Codec na omume
Na-echekwa ma ọ bụ na-ebusa egwu na ụda ihe nke ọma n'otu ụdị nke na-ejikwa ụdị ọdịnaya niile.
Otu dị iche iche na-enwetakarị nsonaazụ ka mma mgbe ha na-akọwapụta ọnụ ụzọ dị mma n'ihu, na-eme ka ụzọ mmadụ si abawanye maka oke ikpe, ma soro ma uru nrụpụta yana ụgwọ njehie n'ime oge.
Ihe ize ndụ & okporo ụzọ nche
Iji olu eme ihe na ihe egwu mpụta ga-abawanye mgbe nkwenye na-efu.
Izi ezi nwere ike ịdaba n'ofe ụda olu, olumba ma ọ bụ gburugburu mkpọtụ.
Enwere ike imehie ọdịyo sịntetik dị ka ezigbo okwu na-enweghị akara doro anya.
Map mmejuputa
Nweta nkwenye doro anya maka ijide olu, imechi, na ijigharị.
Mesoo nke a dị ka ọnụ ụzọ ámá ihe akaebe: ọ bụrụ na emezughị njirisi, kwụsịtụ mbugharị, mechie oghere ahụ, naanị wee gbasaa ojiji.
Nwale ogo n'ofe ndị na-ekwu okwu dị iche iche yana ọnọdụ ndabere.
Mesoo nke a dị ka ọnụ ụzọ ámá ihe akaebe: ọ bụrụ na emezughị njirisi, kwụsịtụ mbugharị, mechie oghere ahụ, naanị wee gbasaa ojiji.
Kọwaa mgbe mmadụ ga-enyocha ma ọ bụ kwado nsonye.
Mesoo nke a dị ka ọnụ ụzọ ámá ihe akaebe: ọ bụrụ na emezughị njirisi, kwụsịtụ mbugharị, mechie oghere ahụ, naanị wee gbasaa ojiji.
Deba aha ọdịyo sịntetik ma debe ndekọ ihe ndekọ maka ịza ajụjụ.
Mesoo nke a dị ka ọnụ ụzọ ámá ihe akaebe: ọ bụrụ na emezughị njirisi, kwụsịtụ mbugharị, mechie oghere ahụ, naanị wee gbasaa ojiji.
Nọgide na-eme nchọpụta
Check your understanding
Test yourself: take the SoundStream Neural Codec quiz