Umbhalo we AI ukuya kuSpeechName

I-Kokoro yimodeli yombhalo-ukuthetha eneparameter ezili-82 ezili-million eyenza ungqubano oluhle ngaphezulu kweqela layo lobunzima. Nangona ubungakanani bayo buncinci, ivelisa ukuthetha okucacileyo nobucacileyo. I-Kokoro ixhasa ulwimi oluninzi oluquka isiNgesi, isiJaphani, isiTshayina, nesiKorea ngeendlela ezahlukeneyo zesandi ezicacileyo. Isebenza ngokukhawuleza kakhulu — ivelisa isandi esimalunga ne-100x ngokukhawuleza kunexesha elibonakalayo kwi-GPU.

I-Piper yinjini elula yombhalo-ukuthetha ephuhliswe yi Rhasspy esebenzisa i VITS kunye ne-larynx architectures. Isebenza ngokupheleleyo kwi CPU, iyenza ibe yindawo efanelekileyo yezixhobo zesiphelo, ulawulo lwasekhaya, kunye neenkqubo ezifuna i-offline TTS. Ngeelizwi ezingaphezu kwe-100 ezisuka kwiilwimi ezingaphezu kwe-30, i-Piper inikezela ngokuthetha okuziva ngathi kuqhelekanga kwisantya sexesha elibonakalayo nakwi-Raspberry Pi 4.

VITS (I-Variation Inference ne-adversarial learning for end-to-end Text-to-Speech) yindlela efana ne-end-to-end TTS evelisa isandi esininzi esiqhelekileyo kunezikhokelo zenqanaba elinye. Isebenzisa i-variation inference ephuculweyo ngokuhamba okuqhelekileyo kunye nenkqubo yoqeqesho oluchaphazelayo, efumana ukuphuculwa okubalulekileyo kwindalo.

MeloTTS yi MyShell. ai yi TTS yelayibrari exhasa isiNgesi (iMelika, iBrithani, i-Indian, i-Australian), isiSpanyol, isiFrentshi, isiTshayina, isiJaphani, nesiKorea. Ikhawuleza kakhulu, iqhubekekisa umbhalo kwisantya esifutshane sexesha elibonakalayo kwi CPU kuphela. MeloTTS icwangciswe ukusetyenziswa kokwenza imveliso kwaye ixhasa zombini i CPU ne GPU inference.

Bark ngu Suno yimodeli yombhalo- ukuya- kwisandi esekelwe kwi-transformer enokuthi ivelise ulwimi oluninzi olunobuntu obuphezulu, kunye nezinye iingoma ezinjengemiculo, ingxolo engaphakathi, kunye neziphumo zesandi. Iyakwazi ukudala unxibelelwano olungathethiweyo njengenkwenkwezi, ukuxoka, nokuxoka. Bark ixhasa ngaphezulu kweendawo ezimiselweyo zesandi ezili-100 kunye neelwimi ezili-13+.

Bark Small yifomati eguqulwe kancinane yemodeli ye Bark ethengisa umgangatho wesandi ngesantya esikhawulezayo sokuzimisela kunye nemfuneko yobume obuphantsi bobume. Igcina ukhono lwe Bark lokuvelisa ulwimi oluneemvakalelo, uxolo, kunye neelwimi ezininzi.

I-CosyVoice 2 yi-Alibaba' s Tongyi Lab ifumana umgangatho wokuthetha othelekiswa nomuntu nge latency ephantsi kakhulu, eyenza ukuba ibe yindawo efanelekileyo yesicelo sexesha elibonakalayo. Isebenzisa indlela yokwahlula i-quantization ye-scalar ephelelayo yokuhambisa uthungelwano kunye noxhasa ukuclonelwa kwelizwi elingekhoyo, uthungelwano lwesiNgesi, kunye nolawulo lweemvakalelo ezincinci. Isebenza kakuhle kunezinye iindlela zentengiso ze-TTS kwiziphumo zovavanyo.

I-Dia yi-Nari Labs yi 1. 6B parameter yombhalo- ukuya- ku- ulwimi lwemodeli eyenziwe ngokukodwa ukudala unxibelelwano lomthumeli- omkhulu. Iyakwazi ukudala unxibelelwano olunombala phakathi kwamathumeli amabini ngokuthatha umjikelo ofanelekileyo, i-prosody, kunye nokubonisa iimvakalelo. I-Dia igqibelele ukudala imixholo yohlobo lwepodcast, unxibelelwano lweencwadi zesandi, kunye ne-AI ethetha- thetha.

I-Parler TTS yimodeli yombhalo-ukuthetha esebenzisa ukuchazwa kwelizwi leelwimi eziqhelekileyo ukulawula ukuthetha okuveliswe. Ngelixa ukhetha ukusuka kwilizwi elimiselweyo, uchaza ilizwi ofuna (umzekelo, "ilizwi lentombazana eshushu enesivakalisi esincinci saseBrithani, ethetha ngokucothayo nocacileyo") kwaye i-Parler ivelisa ukuthetha ohambelana nolwazi. Oku kwenza ukuba ibe yeyona ilula kwisicelo esidala.

GLM- TTS ngu Zhipu AI yinkqubo yombhalo- ukuya- ku- kuthetha eyenziwe kwi Llama architecture ngothelekiso lokuhamba. Ifumana umyinge womonakalo wophawu olusezantsi phakathi kweemodeli ze- TTS ezivulekileyo, okuthetha ukuba ivelisa ukuthetha okuchanekileyo kakhulu. I- GLM- TTS ixhasa isiNgesi neSitshayina ngokuphindaphinda kwelizwi ukusuka kwi 3- 10 yesibini iiseti zesandi.

I-IndexTTS-2 yinkqubo ephambili yombhalo-ukuthetha eyenza kakuhle kwisandi esingena-nto esidityanisiweyo kunye nolawulo lweemvakalelo ezinogranule. Iyakwazi ukudala ukuthetha ngeetoni ezikhethekileyo zeemvakalelo ezinjengeemnandi, ezibuhlungu, ezixhaphakileyo, okanye ezixhalabisayo ngaphandle kokufuna i-data yoqeqesho lweemvakalelo ezikhethekileyo. Imodeli isebenzisa i-emotional vectors ukulawula ngokuchanekileyo ukubonisa kweemvakalelo zelizwi eliveliswe.

I-Spark TTS ngu-SparkAudio yimodeli yombhalo-ukuthetha edibanisa ukuclonelwa kwelizwi ngeemvakalelo ezilawulwayo kunye nesitayile sokuthetha. Ukusebenzisa kuphela imizuzwana emihlanu yoluhlu lwesandi, inokuklona ilizwi kwaye emva koko ivelise ulwimi ngeemvakalelo ezahlukeneyo, isantya, kunye nesitayile ngelixa igcina uqhagamshelwano lwelizwi eliklonwe. I-Spark TTS isebenzisa inkqubo yolawulo olusekwe kwi-prompt.

GPT- SoVITS idibanisa i-GPT-style ulwimi lohlobo kunye ne SoVITS (Ukwahlula ngelizwi ngeNguqulelo kunye neSynthesis) ukufana kwelizwi elinamandla elincinci-lokubetha. Ngemizuzu emihlanu yobhekiso lwesandi, inokufana ngenene nelizwi kwaye ivelise ulwimi olutsha ngelixa igcina iimpawu ezikhethekileyo zomthunywa. Isebenza kakuhle kukuthetha kunye nokudibanisa ngelizwi.

I-Orpheus yimodeli enkulu yombhalo-ukuthetha-ukuthetha efumana ukubonakaliswa kweemvakalelo kwinqanaba lomntu. Iqeqeshwe ngaphezulu kweeyure ezili-100,000 ze data yokuthetha eyahlukeneyo, i excels ekuveliseni ukuthetha ngeemvakalelo eziqhelekileyo, uxinzelelo, kunye neendlela zokuthetha. I-Orpheus inokuvelisa ukuthetha okungaqhelekanga ukusuka kushicilelo lomntu.

Ibhokisi yencoko yababini ngu Resemble AI yimodeli yokuklonya yelizwi eliphambili le-zero-shot. Iyakwazi ukubuyisela nayiphi na ilizwi ukusuka kwisampuli yesandi epheleleyo ngempumelelo ephawulekayo, ithatha hayi kuphela i-timbre kodwa nohlobo lokuthetha kunye nemibala eqaqambileyo. Ibhokisi yencoko yababini ikwabonisa ulawulo lweemvakalelo ezincinci, ezikuvumela ukuba ulungelelanise into eqaqambileyo yelizwi eliveliswe ngokuzimeleyo ukusuka kwilizwi elichaziweyo.

I-Tortoise TTS yinkqubo yokubhala- ukuya- ku- kuthetha enesandi esininzi esiphindayo esinika ingqalelo umgangatho wesandi ngaphezulu kwesantya. Isebenzisa uyilo lwe DALL- E- inspired ukudala ulwimi oluqhelekileyo kakhulu nge- prosody elungileyo kunye nohlobo lomvakalisi. Xa ihamba phantsi kunezinye iindlela ezininzi, i- Tortoise ivelisa ezinye zezinye zezinto ezibonakalayo zelizwi elifumanekayo kwindlela yokusebenza yomthombo ovulekileyo.

I-StyleTTS 2 ifumana uxinzelelo lwe-TTS lomgangatho womntu ngokudibanisa ukusasazeka kwesicwangciso kunye noqeqesho oluchaseneyo lusebenzisa iimodeli ezinkulu zesivakalisi. Ivelisa isivakalisi esidlangalaleni phakathi kweemodeli zomthumeli omnye, esinokhuphisana neengxelo zomntu. I-StyleTTS 2 isebenzisa ukusasazeka-okusekelwe kuyilo lwesivakalisi ukutsala uluhlu olupheleleyo lotshintsho lwesivakalisi somntu.

OpenVoice ngu MyShell. ai ivumela ukuklonya kwelizwi ngokuzenzekelayo ngolawulo olukhulu ngaphezulu kwendlela yelizwi, iimvakalelo, isivakalisi, umculo, izithuba, kunye ne-intonation. Iyakwazi ukuklonya ilizwi ukusuka kwiclip yesandi esifutshane kwaye ivelise ukuthetha kwiilwimi ezininzi ngelixa igcina uphawu lomthumeli. OpenVoice isebenza njengengcambu yelizwi, ivumela ukuguqulwa kwelizwi ngexesha elibonakalayo.

Qwen3- TTS yimodeli yombhalo- ukuya- ku- kuthetha yeparameter eyi- 1. 7 yezigidigidi ukusuka kwiqela le Qwen le Alibaba. Ixhasa iindlela ezintathu: ilizwi elimiselweyo elinomlawuli weemvakalelo (abathethi aba- 9), ukuclonelwa kwelizwi ukusuka kwimizuzu emi- 3 kuphela yesandi, kunye nendlela yoyilo lwelizwi elikhethekileyo apho uchaza khona ilizwi ofuna ngalo kwilwimi oluqhelekileyo. Iquka ulwimi oluli- 10 olunokubonisa okuphezulu kunye ne- prosody eqhelekileyo.

I-Sesame CSM (iModeli yoMbhalo weNtetho) yimodeli eneparameter ezizigidi ezili-1 ezidweliswe ngokukodwa ukuvelisa umbhalo wencoko. Imodeli imilinganiselo eqhelekileyo yencoko yomuntu kubandakanya ukujika-ukuthatha ixesha, uphendule umjelo, uphendule ngeemvakalelo, kunye nokuhamba kwencoko. I-CSM ivelisa umsindo oziva ngathi ngumbhalo wencoko yomuntu oqhelekileyo kunokuba ngumbhalo owenziwe ngesandla.

Ibhokisi yencoko yababini ye Turbo yi Resemble AI yi 350M uhlaziyo lweparameter kwibhokisi yencoko yababini, inikezela ukuya kwisantya sexesha elipheleleyo le 6x nge sub- 200ms latency. Ixhasa ii tags ze paralinguistic ezinjenge [laugh], [cough], kunye ne [chuckle] ngqo kumbhalo. Iquka i Perth watermarking kuzo zonke isandi eziveliswe ukulandela ipropathi.

Zonos v0. 1 yi Zyphra yimodeli yeparameter ye 1. 6B ebonisa ulawulo lweemvakalelo ezinogranule ezincinci kunye neslide zethemba, inkanuko, ubuhlungu, uxinzelelo, kunye nokuxhaphaza. Inikezela zombini iTransformer kunye ne- novel SSM (imodeli yendawo- yendawo) utshintsho. Iqeqeshwe kwi 200K + iiyure zokuthetha ngeelwimi ezininzi nge- zero- shot voice cloning ukusuka kwi 10- 30 imizuzwana yesandi esibhekisa kuyo.

Dia2 yi Nari Labs iqulethe uhlaziyo lwe-streaming-first kwi Dia, efumaneka kwi 1B ne 2B iparameter. Iqala ukudibanisa isandi ukusuka kwitokeni ezimbalwa zokuqala, yenza ukuba ilungele ii-real-time voice agents kunye ne-speech-to-speech pipelines. Ixhasa unxibelelwano lomthumeli-omninzi nge [S1] / [S2] tags kunye ne paralinguistic cues njenge (laughs), (coughs).

VoxCPM 1. 5 yi OpenBMB yimodeli entsha ye TTS engena- tokenizer esebenza kwisithuba esiqhubekayo kunokuba yi- tokens efihlakeleyo. Ivelisa i-audio ye-44. 1kHz ethembekileyo, ixhasa ukuklona kwelizwi elingekhoyo-liqhutywe ukusuka kwimizuzu emi-3- 10, kwaye igcina ukulingana phakathi kweparagraphs. Ukuklona kwe-cross- language kuvumela ukuba usebenzise ilizwi lesiNgesi kwilizwi laseTshayina kwaye ngokuthe ngqo.

OuteTTS iqhuba iimodeli ezinkulu zolwimi ngemisebenzi yokubhala-ukuze-uthethe ngelixa igcina uyilo oluphambili. Ixhasa ii-backends ezininzi kubandakanya i-lama.cpp (CPU/GPU), Ukutsala i-Face Transformers, ExLlamaV2, VLLM, naphi na ukuqonda kwebrowser nge-Transformers.js. Iimpawu zokuklona kwelizwi elingenanto-eyenziweyo ngeeprofayili zomthumeli ezigcinwe njenge-JSON.

I-TADA (i-Text-Acoustic Dual Alignment) nguHume AI yimodeli ye-TTS ephambili esusa iimvakalelo ezingenangqondo ngenkqubo entsha yoyilo lwemimiselo emibini eyenziwe kwi-Llama 3.2. Ifumaneka kwi-1B (isiNgesi) kunye ne-3B (iilwimi ezininzi) iimodeli, i-TADA ifumana i-RTF ye-0.09 — 5x ikhawulezayo kuneemodeli ze-LLM-based TTS ezithelekiswayo. Ixhasa ukuya kuthi ga kwi-700 yeesekondi zemeko yesandi kwaye ivelisa ulwimi olubonisa ububele ngaphandle kweemvakalelo ezingabonakaliyo kwiimvakalelo eziqhelekileyo.

I VibeVoice ye Microsoft ivela ngeendlela ezimbini: imodeli ye 1. 5B yezinto eziqulethe iifomu ezide (zifike kwimizuzu engama- 90, abavakalisi aba- 4) kunye nexesha elipheleleyo lemodeli ye 0. 5B yosasazo nge ~200ms yokuqala yokungasebenzi kakuhle kwesandi. I 1. 5B ilungele iipodcasts kunye neencwadi zesandi ezinesandi esizinzileyo ngaphezulu kweendawo ezide. Qaphela: I Microsoft isuse ikhowudi ye TTS kwindawo yokugcina kwaye ivelise isandi equlethe i- AI ekwazi ukuvakala.

I Pocket TTS ngu Kyutai (abavelisi be Moshi) yimodeli yombhalo- ukuya- ku- kuthetha encinci eneparameter ye 100M eyenza ubunzima bayo. Isebenza kakuhle kwi CPU, ixhasa ukuklona kwesandi esingenanto ukusuka kwisampuli yesandi, kwaye ivelisa ulwimi oluzimeleyo. Ubungakanani bemodeli encinci yenza ukuba ibe yindawo efanelekileyo yokubekwa kwesiphelo kunye nemeko- bume ephantsi yecebo.

Kitten TTS by KittenML is an ultra-lightweight text-to-speech model built on ONNX. With variants from 15M to 80M parameters (25-80 MB on disk), it delivers high-quality voice synthesis on CPU without requiring a GPU. Features 8 built-in voices, adjustable speech speed, and built-in text preprocessing for numbers, currencies, and units. Ideal for edge deployment and low-latency applications.

CosyVoice3 is the latest evolution from Alibaba's FunAudioLLM team. It features bi-streaming inference with ~150ms latency, instruction-based control for emotion/speed/volume, and improved speaker similarity for zero-shot cloning. Supports 9 languages plus 18 Chinese dialects. RL-tuned variant delivers state-of-the-art prosody.

MOSS-TTS from OpenMOSS supports generation of up to 1 hour of continuous speech across 20 languages. Features token-level duration control, phoneme-level pronunciation control via IPA/Pinyin, and code-switching between languages. The 8B production model delivers state-of-the-art quality with zero-shot voice cloning from reference audio.

MegaTTS3 from ByteDance uses a novel sparse alignment mechanism combined with a latent diffusion transformer. Features adjustable trade-off between speech intelligibility and speaker similarity for zero-shot voice cloning.