AI Text to SpeechName

Kokoro ndi 82 miliyoni paramita malemba-ku-kulankhula chitsanzo chomwe punches bwino pamwamba pa khalidwe lake la thupi. Ngakhale ndi ochepa kukula, amatulutsa mawu owoneka bwino ndi owoneka bwino. Kokoro amathandiza mabungwe ambiri kuphatikizapo Chijeremani, Chijeremani, Chijeremani, ndi Korean ndi mitundu yosiyanasiyana ya mawu owoneka bwino.

Piper ndi makina otsika mtengo a mawu ochokera ku mawu omwe adapangidwa ndi Rhasspy omwe amagwiritsa ntchito VITS ndi larynx architectures. Imayenda kwathunthu pa CPU, zomwe zimapangitsa kuti ikhale yabwino kwa zida za edge, zowongolera zanyumba, ndi mapulogalamu omwe akufuna TTS osagwirizana. Ndi mawu oposa 100 m'zinenero 30 +, Piper imabweretsa mawu owoneka bwino panthawi ya real-time ngakhale pa Raspberry Pi 4.

VITS (Variational Inference with adversarial learning for end-to-end Text-to-Speech) ndi njira yofanana yoyambira kumapeto kwa TTS yomwe imapanga mawu owoneka bwino kwambiri kuposa mamodeli anthawi zonse awiri. Imagwiritsa ntchito kutengera kwa maonekedwe osiyanasiyana omwe amawonjezeredwa ndi kuwongolera kwa magazi ndi njira yophunzitsa yotsutsana, yomwe imakwaniritsa kuwonjezeka kwakukulu kwa chilengedwe.

MeloTTS ndi MyShell.ai ndi TTS library yokhala ndi mabuku ambiri omwe amathandizira Chijeremani (cha America, cha British, cha Indian, cha Australia), Chisipanishi, Chifalansa, Chijeremani, cha Japanese, ndi cha Korean. Ndiyabwino kwambiri, yopanga malemba panthawi yoyenera kwambiri pa CPU yokha. MeloTTS idapangidwa kuti igwiritse ntchito kupanga ndipo imathandizira kuzindikira kwa CPU ndi GPU.

Bark ndi Suno ndi transformer-ogwirizana malemba-ku-audio model kuti angayambitse kwambiri realistic, mawu ambirimbiri monganso ena audio monga nyimbo, fumbi m'mbuyo, ndi zotsatira za mawu.It angayambitse nonverbal kulankhulana monga kuseka, kupweteka, ndi kugwa. Bark amathandiza zoposa 100 speaker presets ndi 13 + zinenero.

Bark Small ndi mtundu wosinthidwa wa Bark model womwe umagulitsa zina mwa zinthu za audio kuti zikhale ndi kuthekera kofulumira kwambiri komanso zofunikira za kukumbukira.Izi zimasunga kulimba kwa Bark kuti zikhale ndi mawu ndi maganizo, kuseka, ndi mabungwe ambiri.

CosyVoice 2 ya Tongyi Lab ya Alibaba imakwaniritsa mtundu wa mawu ogwirizana ndi munthu ndi latency yayikulu kwambiri, zomwe zimapangitsa kuti zikhale zosavuta kugwiritsa ntchito nthawi yambiri. Imagwiritsa ntchito njira yomaliza ya quantum kwa synthesization ya streaming ndipo imathandizira kujambula mawu opanda malire, synthesization yosiyanasiyana ya zinenero, komanso kuwongolera maganizo olimba.

Dia ndi Nari Labs ndi 1.6B paramita malemba-ku-kulankhula model anapanga mosamala kuti atenge mauthenga ambiri-wolankhula. Ikhoza kupanga zokambirana zowoneka bwino pakati pa olankhula awiri ndi zoyenera turn-kutenga, prosody, ndi kusonyeza chisoni. Dia ndi yabwino kwa kulenga podcast-style zinthu, audiobook mauthenga, ndi interactional conversational AI.

Parler TTS ndi mtundu wa mawu-ku-mawu womwe umagwiritsa ntchito mawu ofotokoza mawu kuti ayang'ane mawu omwe amapangidwa. M'malo mosankha kuchokera ku mawu omwe amakonzedwa, mumafotokoza mawu omwe mukufuna (mwachitsanzo, "mawu a mkazi wotentha ndi mawu a British, akulankhula mofulumira komanso mosawoneka bwino") ndipo Parler imapanga mawu omwe amagwirizana ndi mawu omwe amafotokozedwa. Izi zimapangitsa kuti ikhale yosinthika kwambiri kwa ntchito zopanga.

GLM-TTS ndi pulogalamu yosintha mawu kukhala mawu yomwe imagwiritsa ntchito zida za Llama ndi njira yosagwirizana ya flow matching. GLM-TTS imathandiza Chijeremani ndi Chisipanishi ndi kufalitsa mawu kuchokera pa 3-10 masekondi a mawu.

IndexTTS-2 ndi njira yosinthira mawu kukhala mawu yomwe imagwira ntchito bwino kwambiri pakupanga mawu opanda kanthu ndi kuwongolera maganizo mosasamala kanthu. Ikhoza kupanga mawu ndi maganizo osiyanasiyana monga osangalala, okhumudwa, okhumudwa kapena okhumudwa popanda kufunikira kuphunzira mawu osiyanasiyana.

Spark TTS ya SparkAudio ndi mtundu wa mawu-ku-mawu womwe umaphatikizapo kufalitsa mawu ndi kuwongolera maganizo ndi mtundu wa mawu. Pogwiritsa ntchito masekondi 5 okha a ma audio, imatha kufalitsa mawu ndipo imapanga mawu ndi maganizo, magwiridwe antchito ndi mitundu yosiyanasiyana, popeza imasunga chidziwitso cha mawu chofalitsidwa. Spark TTS imagwiritsa ntchito njira yoyang'anira yomwe imayang'aniridwa ndi uthenga.

GPT-SoVITS imaphatikizapo kujambula kwa mawu a GPT ndi SoVITS (Singing Voice Inference via Translation and Synthesis) kwa kujambula mawu kwamphamvu kwa ma shots ochepa. Ndi 5 masekondi ochepa a audio yolemba, imatha kujambula mawu moyenera ndikupanga mawu atsopano poteteza mfundo zosiyanasiyana za wokamba.

Orpheus ndi mtundu waukulu wa kulemba-ku-kulankhula womwe umakwaniritsa chidziwitso cha chisoni pamalingaliro a munthu. Kuphunzira pa maola oposa 100,000 a deta yosiyanasiyana ya mawu, imagwira bwino ntchito popanga mawu ndi chisoni chachilengedwe, kutanthauzira, ndi mitundu yolankhula.

Chatterbox ndi Resemble AI ndi njira yokhayo yokhayo yokhayo yokhayo yokhayo yomwe imatha kujambula mawu onse kuchokera ku sample imodzi ya audio ndi kuzindikira kwakukulu, kujambula sikuti timbre, komanso mtundu wa mawu ndi ma nuances a maganizo. Chatterbox imaphatikizaponso kuwongolera kwa maganizo oyera, kulola kuti musinthe tonse ya maganizo a mawu omwe amapangidwa mosiyana ndi dzina la mawu.

Tortoise TTS ndi dongosolo losintha mawu kukhala mawu lomwe limagwiritsa ntchito maudindo osiyanasiyana a mawu kuti lipange mawu owoneka bwino kwambiri ndi kugwirizana kwa mawu ndi mawu. Ngakhale kuti ndi lachidule kuposa njira zina zambiri, Tortoise TTS imapanga mawu owoneka bwino kwambiri omwe ali m'makompyuta opanda zingwe.

StyleTTS 2 imakwaniritsa TTS synthesization ya munthu-level pogwiritsa ntchito kulumikizana kwa mtundu ndi kuphunzitsa kwa otsutsana pogwiritsa ntchito mapangidwe a mawu akuluakulu. Amapanga mawu owoneka bwino kwambiri pakati pa mapangidwe a mawu a munthu mmodzi, opikisana ndi zolemba za munthu.

OpenVoice ya MyShell.ai imalola kujambula mawu mwamsanga ndi kuwongolera kwa granular pamtundu wa mawu, chisoni, kuyankhula, ritmu, kuletsa, ndi kuyankhula. Itha kujambula mawu kuchokera ku audio clip yafupi ndi kutulutsa mawu m'zinenero zambiri poteteza chidziwitso cha wolankhula. OpenVoice imagwiranso ntchito ngati wosintha mawu, kulola kusintha kwa mawu panthawi ya real-time.

Qwen3-TTS ndi 1.7 biliyoni paramita malemba-ku-kulankhula mtundu kuchokera Qwen Alibaba's timu. It supports atatu modes: preset mawu ndi kuwongolera maganizo (9 okamba), mawu kloning kuchokera 3 masekondi okha a audio, ndi khalidwe lapadera mawu design mode kumene inu kufotokoza mawu mukufuna mu chilankhulo chachilengedwe.It amaphatikizapo 10 mabungwe a zinenero ndi khalidwe labwino ndi zomvetsera zachilengedwe.

Sesame CSM (Conversational Speech Model) ndi 1 billion parameter model yopangidwa mwapadera kuti ipange uthenga wolankhulana. Imakhala ndi mapangidwe achilengedwe a uthenga wa munthu kuphatikizapo nthawi yosintha, mayankho a backchannel, mayankho a maganizo, ndi uthenga wolankhulana. CSM imapanga audio yomwe imawoneka ngati uthenga wa munthu wachilengedwe kuposa uthenga wopangidwa.

Chatterbox Turbo ndi Resemble AI ndi 350M paramita kuwonjezera kwa Chatterbox, kubweretsa mpaka 6x real-time galimoto ndi sub-200ms latency. It supports paralinguistic tags monga [laugh], [kugwa], ndi [kucheka] mwachindunji mu malemba.

Zonos v0.1 ndi Zyphra ndi 1.6B parameter model yomwe imaphatikizapo kuwongolera kwa maganizo oyera ndi sliders a chisangalalo, nkhawa, kukhumudwa, kukhumudwa, ndi kukhumudwa. Imapatsa Transformer ndi mtundu watsopano wa SSM (modelu ya state-space). Yatsirizidwa pa 200K + maola a mawu ambiri a zinenero ndi kujambula mawu opanda malire kuchokera ku 10-30 masekondi a audio reference.

Dia2 ndi Nari Labs ndi kusinthidwa kwatsopano kwa Dia, kupezeka m'maparamita 1B ndi 2B. Amayamba kutulutsa mawu kuchokera pazolemba zazikuluzikulu, zomwe zimapangitsa kuti zikhale zoyenera kwa olemba mawu anthawi yeniyeni ndi ma pipelines a mawu-ku-mawu. Amathandiza macheza a omvera ambiri ndi [S1] / [S2] tags ndi paralinguistic cues monga (kuchepa), (kupweteka).

VoxCPM 1. 5 ndi OpenBMB ndi mndandanda watsopano wa TTS wopanda tokenizer womwe umagwira ntchito m'malo opitilira muyeso kuposa ma tokens osiyana. Amapanga 44. 1kHz olimba kwambiri, amathandizira kujambula mawu osatha kuchokera pamasekondi 3- 10, ndipo amasunga kulimba pakati pa masamba. Kujambula mawu osiyanasiyana kumakupatsani mwayi wogwiritsa ntchito mawu a Chingelezi kulankhula kwa Chisipanishi ndi vice versa.

OuteTTS imawonjezera mapangidwe amtundu waukulu ndi zofunikira za mawu-ku-mawu poteteza mapangidwe ake oyambirira. Imathandizira ma backends ambiri kuphatikiza llama.cpp (CPU / GPU), Hugging Face Transformers, ExLlamaV2, VLLM, komanso kuzindikira kwa msakatuli pogwiritsa ntchito Transformers.js.

TADA (Text-Acoustic Dual Alignment) ndi Hume AI ndi mtundu wa TTS woyamba womwe umachotsa ma halucinations pogwiritsa ntchito kapangidwe katsopano ka ma alignment awiri opangidwa pa Llama 3.2. Kupezeka mu 1B (Chingelezi) ndi 3B (mawu ambiri), TADA imakwaniritsa RTF ya 0.09 - 5x mofulumira kuposa ma TTS ogwirizana ndi LLM. Imathandizira mpaka masekondi 700 a audio context ndi kutulutsa mawu owoneka bwino ndi zero hallucinations pazofunikira zazikulu.

VibeVoice ya Microsoft imapezeka m’mapangidwe awiri: 1.5B model ya mavidiyo anthawi yayitali (kufika maminitsi 90, okamba 4) ndi Realtime 0.5B model ya mavidiyo oyenda ndi nthawi yoyamba yokhala ndi mawu yopitilira 200ms. 1.5B model imagwira ntchito bwino kwambiri pa podcasts ndi mabuku olankhula ndi mawu ogwirizana panthawi yaitali.

Pocket TTS ya Kyutai (opanga Moshi) ndi mtundu wa 100M wa maparamita a masamba omwe amagwira bwino ntchito. Imagwira ntchito bwino pa CPU, imathandizira kujambula mawu osatha kuchokera pa sampling ya audio imodzi, ndipo imatulutsa mawu owoneka bwino.

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.