Qoraalka AI si uu u noqdo hadal

Kokoro waa 82 milyan oo parameter qoraal-to-hadalka oo qaab ah oo si fiican u saaran heerka miisaanka. In kasta oo ay yar tahay, waxay soo saartaa hadal aad u dabiici ah oo muujinaya. Kokoro waxay taageertaa luqado badan oo ay ku jiraan Ingiriisiga, Japan, Shiinaha, iyo Korean oo leh codyo muujinaya oo kala duwan. Waxay ku socotaa si aad u dhakhso badan - abuurista maqalka oo ku dhow 100x ka dhaqso badan waqtiga dhabta ah ee GPU.

Piper waa mashiin qoraal-u-hadalka ah oo fudud oo ay soo saartay Rhasspy oo isticmaalaya VITS iyo dhismayaasha larynx. Waxay ku socotaa oo dhan CPU, taasoo ka dhigaysa mid aad u fiican qalabka edge, otomaatiga guriga, iyo codsiyada u baahan TTS offline. Iyadoo ku saabsan 100 codadka oo ka socda 30 + luqadood, Piper wuxuu siiyaa hadalka dabiiciga ah ee ku jira xawaaraha waqtiga dhabta ah, xitaa Raspberry Pi 4.

VITS (Isbarbardhiga kala duwan ee la barashada adversarial ee dhamaadka-to-dhamaadka Text-to-Speech) waa isku mid ah dhamaadka-to-dhamaadka TTS habka oo soo saara audio maqal badan oo dabiici ah ka badan hadda laba-geesoodka ah. Waxay qaadataa isbarbardhiga kala duwan oo la kordhiyay la isku mid ah qulqulka iyo habka tababarka adversarial, gaarista horumar weyn oo dabiiciga ah.

MeloTTS by MyShell.ai waa maktabad TTS oo luqado badan leh oo taageera Ingiriisiga (Amerika, Ingiriis, Hindi, Australia), Isbaanish, Faransiis, Shiinaha, Japan, iyo Korean. Waa mid aad u dhaqso badan, oo qoraalka u qaabeeya xawaare ku dhow waqtiga dhabta ah ee CPU keliya. MeloTTS waxaa loogu talagalay isticmaalka wax soo saarka wuxuuna taageeraa labadaba CPU iyo GPU.

Bark by Suno waa transformer-ku salaysan qoraalka-to-audio qaabka oo ay ku abuuri karaan si aad u dhab ah, hadalka luqado badan iyo sidoo kale audio kale sida muusiga, codka background, iyo saamaynta codka. Waxay soo saari kartaa xiriirada aan hadalka ahayn sida qosli, oohin, iyo oohin. Bark taageeraa in ka badan 100 speaker presets iyo 13 + luqadood.

Bark Small waa nooc ka mid ah qaabka Bark oo ah nooc ka mid ah tayada maqalka ah oo si weyn u dhaqdhaqaaqa xawaaraha xawaaraha iyo shuruudaha xusuusta hooseeya. Waxay haystaan ​​qaababka Bark ee lagu abuuro hadalka oo leh dareenka, qoslka, iyo luqadaha badan.

CosyVoice 2 by Alibaba ee Tongyi Lab gaadhay dadka-la barbar dhigi karo tayada hadalka la latentity aad u hooseeya, taas oo ka dhigaysa mid aad u fiican oo loogu talagalay codsiyada waqtiga dhabta ah. Waxay isticmaalaan habka finite scalar quantumization u streaming isku dhafka iyo taageertaa zero-shot codka isku dhafka, cross-afka isku dhafka, iyo fine-grained emotion control.

Dia by Nari Labs waa 1.6B parameter qoraal-to-speech qaabka loogu talagalay gaar ahaan si loo abuuro wada hadalka multi-hoosaad. Waxay soo saari kartaa wada hadalka dabiiciga ah ee dhex mara labada hadalle la wareegidda habboon, prosody, iyo muujinta dareenka. Dia waa ku habboon abuurista podcast-style content, audiobook wada hadalka, iyo AI wada hadalka is dhexgalka.

Parler TTS waa qoraal-in-hadalka oo qaab ah oo isticmaalaya qoraalka afka dabiiciga ah ee codka si ay u xakameeyaan hadalka soo baxay. In kastoo dooranaya codadka hore loo dhigay, waxaad sharaxaysaa codka aad rabto (tusaale ahaan, "hees dumar ah oo kulul oo leh afka Ingiriisiga yar, hadalka oo si degdeg ah oo cad u hadlaya") iyo Parler abuurtaa hadalka la mid ah in sharaxaadda. Tani waxay ka dhigaysaa mid gaar ah oo la jaanqaadi kara codsiyada abuurista.

GLM-TTS by Zhipu AI waa qoraal-to-dhaqanka nidaamka ku dhisan dhismaha Llama la daadashada isku mid ah. Waxay gaadhay xaddiga ugu hooseeya ee xadgudubka xarafka ka mid ah qaababka TTS ee furan, taas oo macnaheedu yahay inay soo saarto afka ugu saxsan. GLM-TTS taageeraa Ingiriisi iyo Shiinaha oo leh codka ka soo 3-10 second audio tusaale.

IndexTTS-2 waa nidaamka qoraalka-to-hadalka oo horumarsan oo ku fiican in aan la maqal-shot codka isku-dhafan oo leh xakamaynta dareenka fine-grained. Waxay abuuri kartaa hadalka la mid ah tonos gaar ah ee dareenka sida farxad, murugada, cadhada, ama cabsi leh aan u baahnayn xogta tababarka dareenka gaar ah.

Spark TTS by SparkAudio waa mid qoraal-to-hadalka oo qaab ah oo isku dhafan codka isku dhafan oo leh dareenka la xakameyn karo iyo qaabka hadalka. Iyadoo la adeegsanayo kaliya 5 ilbiriqsi oo ka mid ah raadinta audio, waxay awood u leedahay in ay isku dhafan codka ka dibna abuuro hadalka la dareenka kala duwan, xawaaraha, iyo qaababka iyadoo la ilaalinayo aqoonsiga codka la isku dhafan. Spark TTS isticmaalaa nidaamka xakamaynta prompt-ku salaysan.

GPT-SoVITS isku daraa GPT-style luqadda qaabeynta la SoVITS (Singing Voice Inference via Translation iyo Synthesis) si ay awood badan qaar ka mid ah-shot codka isku-dhafan. Iyadoo yar sida 5 ilbiriqsi oo ah tilmaamaha audio, waxay si sax ah u isku-dhafan codka iyo abuuro hadal cusub oo ay ilaaliyaan astaamaha gaarka ah ee hadalka.

Orpheus waa mid aad u ballaaran oo qoraalka-to-hadalka tusaale ah in uu gaaro hadalka dareenka heerka aadanaha. Tababar ku saabsan 100,000 saacadood oo xogta hadalka kala duwan, wuxuu ku fiican yahay in uu abuuro hadalka la dareenka dabiiciga ah, xoogga, iyo qaababka hadalka. Orpheus waxaa laga yaabaa in ay soo saaraan hadalka in uu yahay virtual indistinguishable ka diiwaangelinta aadanaha.

Chatterbox by Resemble AI waa mid ka mid ah ugu dambeeyay zero-shot codka isku-dhafan qaabka. Waxay ku soo celin kartaa cod kasta oo ka mid ah tusaale audio kaliya oo leh saxnaanta la yaab leh, oo aan kaliya timbre laakiin sidoo kale qaabka hadalka iyo nuances dareenka. Chatterbox sidoo kale waxaa ka mid ah fine-grained dareenka xakamaynta, kuu ogolaanaya inaad hagaajiso qaabka dareenka hadalka soo saara madaxbannaan ka aqoonsiga codka.

Tortoise TTS waa autoregressive multi-codad qoraalka-to-dhageysiga nidaamka in ay hormuudka u tahay tayada audio ka badan xawaaraha. Waxay isticmaalaan DALL-E-inspired dhisme si ay u abuuraan hadalka aad u dabiici ah oo leh prosody wanaagsan iyo isku midnimada afhayeenka. Iyadoo ka badnaan badan oo badan oo ka baxsan, Tortoise soo saarta qaar ka mid ah hadalka ugu dhabta ah ee isku dhafan ee ku jira open-source ecosystem.

StyleTTS 2 gaadhay TTS heerka aadanaha-soo-saarka by isku darka qaabka faafinta la tababarka adversarial isticmaalaya qaabab afka weyn. Waxay soo saartaa hadalka ugu dabiiciga ah ee dhexdhexaad ah oo ka mid ah qaabab hal-hoogaamiyaha, riixaya diiwaangelinta aadanaha. StyleTTS 2 isticmaalaa faafinta-ku salaysan qaabka qaabeynta si ay u qabtaan kala duwanaansho hadalka aadanaha oo dhan.

OpenVoice by MyShell.ai u oggolaanaysaa codka instant isku-dhafan oo leh xakamaynta granular ee qaabka codka, dareenka, afka, riiq, joojinta, iyo intonation. Waxay ka heli kartaa cod ka mid ah audio clip gaaban oo abuuro hadalka luqado badan oo ay ilaaliyaan aqoonsiga hadalka. OpenVoice sidoo kale waxay u shaqeysaa sida codka ku beddele, u oggolaanaya isbeddelka codka waqtiga dhabta ah.

Qwen3-TTS waa 1.7 bilyan parameter qoraal-to-hadalka qaabka ka Alibaba's Qwen koox. Waxay taageertaa saddex qaabab: codadka hore loo dhigay oo leh xakamaynta dareenka (9 hadal jeediyayaashii), codka isku-dhafan oo ka mid ah kaliya 3 ilbiriqsi oo audio, iyo qaabab muuqaal ah oo codka ah oo aad ku sharaxdo codka aad rabto in aad afka dabiiciga ah.

Sesame CSM (Conversational Speech Model) waa 1 bilyan parameter tusaale loogu talagalay gaar ahaan abuurista hadalka hadalka. Waxay u egtahay qaababka dabiiciga ah ee hadalka aadanaha oo ay ku jiraan wareega-qaado waqti, jawaabaha backchannel, falcelinta dareenka, iyo qulqulka hadalka. CSM soo saarta audio oo u eg hadalka dabiiciga ah ee aadanaha ka badan hadalka synthesized.

Chatterbox Turbo by Resemble AI waa 350M parameter kor u qaadista Chatterbox, soo bandhigaya ilaa 6x xawaare waqtiga dhabta ah oo leh sub-200ms latentity. Waxay taageertaa tags paralinguistic sida [qas,] [cadho], iyo [qas] si toos ah qoraalka.

Zonos v0.1 by Zyphra waa 1.6B parameter tusaale ah oo leh fine-grained dareenka xakamaynta sliders u farxad, cadhada, murugada, cabsida, iyo yaab. Waxay bixisaa labadaba Transformer iyo sheeko SSM (state-space tusaale) nooc. Tababar ku 200K + saacadood oo hadalka luqado badan la zero-shot codka ka soo koobista 10-30 ilbiriqsi ee tilmaame audio.

Dia2 by Nari Labs waa dib u eegis-hore u ah Dia, oo laga heli karo noocyada 1B iyo 2B. Waxay bilaabeysaa inay ka soo saarto codka ka hor inta aan la bilaabin, taas oo ka dhigaysa mid ku habboon wakiilada codka iyo xarumaha hadalka-ugu-hadalka. Waxay taageertaa wada hadalka multi-hoosaad leh [S1] / [S2] tags iyo calaamadaha paralinguistic sida (qasuuqa), (caabuqa).

VoxCPM 1.5 by OpenBMB waa nooc cusub oo TTS ah oo aan lahayn tokenizer oo ka shaqeeya meel joogto ah oo aan ahayn tokens kala duwan. Waxay soo saartaa 44.1kHz audio oo tayo sare leh, waxay taageertaa isku-dhafka codka zero-shot ee 3-10 ilbiriqsi, waxayna ilaalisaa isku-dhafka qaybaha.

OuteTTS wuxuu ballaariyaa moodooyinka afka oo leh awoodo qoraal-u-hadalka ah, isagoo ilaalinaya naqshadeynta asalka ah. Waxay taageertaa backends badan oo ay ku jiraan llama.cpp (CPU / GPU), Hugging Face Transformers, ExLlamaV2, VLLM, iyo xitaa soo jeedinta brauzer-ka ee Transformers.js.

TADA (Text-Acoustic Dual Alignment) by Hume AI waa qaab TTS ah oo ka saara hal-abuurka hal-abuurka ah ee ku saleysan Llama 3.2. Waxaa laga heli karaa 1B (Ingriis) iyo 3B (farsamo badan) noocyada, TADA wuxuu gaarayaa RTF ah 0.09 - 5x ka dhaqso badan oo ka mid ah LLM-ku saleysan TTS-yada la midka ah. Waxay taageertaa ilaa 700 ilbiriqsi oo ah kontekst audio iyo waxay soo saartaa hadallo dareenka leh oo leh hal-abuurro aan la aqoon.

VibeVoice by Microsoft waxaa ka mid ah laba nooc: a 1.5B qaabka content dheer-form (ugu badnaan 90 daqiiqo, 4 hadal jeediyay) iyo Realtime 0.5B qaabka u streaming la ~ 200ms hore audio latency. The 1.5B nooc ka fiican podcasts iyo buugaagta la hadalka la isku mid ah inta badan wadooyinka dheer.

Pocket TTS by Kyutai (abuurayaasha Moshi) waa 100M parameter qoraalka-to-speech qaabka compact in ka badan miisaanka. Waxay ku socotaa si wax ku ool ah on CPU, taageertaa zero-shot codka ka soo baxa tusaale audio kaliya, iyo soo saarta hadalka dabiiciga ah-soo baxay.

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.