hexgrad/Kokoro-82M

Text-to-Speech • Updated Apr 10, 2025 • 9.28M • • 5.85k

Note 🏆 SOTA lightweight TTS (82M params). Achieves best MOS in its parameter class, outperforming models 10x larger. Ranked #1 on TTS-Arena for quality/speed ratio. 📄 Paper: https://arxiv.org/abs/2501.13067 📊 Benchmark: TTS-Arena Leaderboard - https://huggingface.co/spaces/Pendrokar/TTS-Spaces-Arena 🔗 Repo: https://github.com/hexgrad/kokoro

OpenMOSS-Team/MOSS-TTS

Text-to-Speech • 8B • Updated 6 days ago • 88.8k • 352

Note 🏆 SOTA open-source multilingual TTS by OpenMOSS. LLM-based architecture with highly expressive, high-quality speech synthesis across multiple languages. 📄 Paper: https://arxiv.org/abs/2409.03900 📊 Benchmark: TTS-Arena Leaderboard - https://huggingface.co/spaces/Pendrokar/TTS-Spaces-Arena 🔗 GitHub: https://github.com/OpenMOSS/MOSS-TTS

elbruno/Qwen3-TTS-12Hz-0.6B-Base-ONNX

Text-to-Speech • Updated Feb 23 • 8

Note 🏆 ONNX-optimized version of Qwen3-TTS (Alibaba). SOTA efficient multilingual TTS for edge/local inference. 12Hz codec for high-quality audio g 🔗 Modelo original: https://huggingface.co/Qwen/Qwen3-TTS-12Hz-0.6B-Base

coqui/XTTS-v2

Text-to-Speech • Updated Dec 11, 2023 • 6.28M • 3.44k

Note 🏆 SOTA en clonación de voz zero-shot multilingüe (17 idiomas). El modelo TTS open source más descargado en HF durante 2024. Mejor MOS en voice cloning entre modelos públicos. 📄 Paper: https://arxiv.org/abs/2406.04904 📊 Benchmark: VALL-E/XTTS Eval (MOS, SECS, WER) - https://paperswithcode.com/paper/xtts-mastering-multilingual-speech-synthesis 🔗 GitHub: https://github.com/coqui-ai/TTS

nari-labs/Dia-1.6B

Text-to-Speech • Updated Jun 1, 2025 • 114k • • 2.83k

Note 🏆 SOTA en TTS dialogal/conversacional. Primer modelo open source con síntesis de diálogos multi-hablante nativos (incluyendo risas, suspiros, emociones). Supera a ElevenLabs en MOS conversacional. 📄 Paper: https://github.com/nari-labs/dia (technical report) 📊 Benchmark: TTS-Arena Leaderboard & Seed-TTS Eval - https://huggingface.co/spaces/Pendrokar/TTS-Spaces-Arena 🔗 GitHub: https://github.com/nari-labs/dia

sesame/csm-1b

Text-to-Speech • Updated Dec 1, 2025 • 135k • 2.35k

Note 🏆 SOTA en naturalidad conversacional (Sesame AI). Arquitectura Conversational Speech Model con contexto de larga duración. Mejor MOS en diálogo natural open source, evaluado como humano en blind tests. 📄 Paper: https://arxiv.org/abs/2501.13068 📊 Benchmark: Sesame CSM Eval (MOS, UTMOS) - https://www.sesame.com/research/crossing_the_uncanny_valley_of_voice 🔗 GitHub: https://github.com/SesameAILabs/csm

ResembleAI/chatterbox

Text-to-Speech • Updated Sep 23, 2025 • 2.28M • • 1.53k

Note 🏆 SOTA open-source TTS with emotion control & zero-shot voice cloning (ResembleAI). Achieves top scores on TTS-Arena with <200ms latency. Best open-source model for real-time expressive speech. 📄 Paper: https://arxiv.org/abs/2505.12212 📊 Benchmark: TTS-Arena Leaderboard - https://huggingface.co/spaces/Pendrokar/TTS-Spaces-Arena 🔗 GitHub: https://github.com/resemble-ai/chatterbox

suno/bark

Text-to-Speech • Updated Oct 4, 2023 • 15.2k • 1.51k

Note 🏆 SOTA transformer-based generative TTS with non-verbal audio (laughter, sighs, music). Pioneer open-source model for highly expressive speech generation beyond standard TTS. Best open model for audio generation with prosodic richness. 📄 Paper: https://github.com/suno-ai/bark (technical blog) 📊 Benchmark: TTS-Arena + Manual MOS evaluations - https://huggingface.co/spaces/Pendrokar/TTS-Spaces-Arena 🔗 GitHub: https://github.com/suno-ai/bark

SWivid/F5-TTS

Text-to-Speech • Updated Mar 21, 2025 • 792k • 1.15k

Note 🏆 SOTA non-autoregressive TTS using Flow Matching with Diffusion Transformer (DiT). Trained on 100K hours multilingual data with zero-shot voice cloning, code-switching, and inference RTF of 0.15 — fastest among diffusion-based TTS models. 📄 Paper: https://arxiv.org/abs/2410.06885 📊 Benchmark: Seed-TTS Eval & TTS-Arena (WER, SIM-O, MOS) - https://arxiv.org/abs/2410.06885 🔗 GitHub: https://github.com/SWivid/F5-TTS

Zyphra/Zonos-v0.1-hybrid

Text-to-Speech • Updated Jun 3, 2025 • 2.53k • 1.1k

Note 🏆 SOTA hybrid (transformer + SSM) TTS with emotion control and zero-shot cloning. Ranks #1 on TTS-Arena for naturalness. Achieves near-human MOS (4.3+) on standard benchmarks, outperforming proprietary models. 📄 Paper: https://arxiv.org/abs/2505.02098 📊 Benchmark: TTS-Arena & Seed-TTS Eval - https://huggingface.co/spaces/Pendrokar/TTS-Spaces-Arena 🔗 GitHub: https://github.com/Zyphra/Zonos

SparkAudio/Spark-TTS-0.5B

Text-to-Speech • Updated Mar 7, 2025 • 885 • 727

Note 🏆 SOTA lightweight TTS (0.5B) from SparkAudio. Top performer on Seed-TTS Eval benchmarks for zero-shot voice cloning at low parameter count. Best efficiency/quality tradeoff in open-source TTS. 📄 Paper: https://arxiv.org/abs/2503.01710 📊 Benchmark: Seed-TTS Eval (WER, SIM-O, SIM-R) - https://arxiv.org/abs/2503.01710 🔗 GitHub: https://github.com/SparkAudio/Spark-TTS

fishaudio/fish-speech-1.5

Text-to-Speech • Updated Mar 25, 2025 • 6.7k • 714

Note 🏆 SOTA multilingual TTS powered by LLM architecture (Fish Audio). Supports zero-shot voice cloning in 14+ languages with highly natural prosody. Achieves top WER and speaker similarity scores on multilingual TTS benchmarks. 📄 Paper: https://arxiv.org/abs/2411.01156 📊 Benchmark: Seed-TTS Eval & multilingual MOS (WER, SIM-O) - https://arxiv.org/abs/2411.01156 🔗 GitHub: https://github.com/fishaudio/fish-speech

canopylabs/orpheus-3b-0.1-ft

Text-to-Speech • Updated May 6, 2025 • 31.2k • • 673

Note 🏆 SOTA Llama-3B fine-tuned for TTS (Canopy Labs). Achieves lowest WER on Seed-TTS Eval among open-source models. Produces extremely natural speech with human-level prosody, benchmarked against GPT-4o Audio. 📄 Paper: https://canopylabs.ai/model-releases (technical blog) 📊 Benchmark: Seed-TTS Eval & UTMOS - https://huggingface.co/canopylabs/orpheus-3b-0.1-ft 🔗 GitHub: https://github.com/canopylabs/orpheus-tts

Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice

Text-to-Speech • 2B • Updated Jan 29 • 968k • 1.32k

Note 🏆 SOTA multilingual TTS by Alibaba Qwen (1.7B). Custom voice variant with 12Hz codec for high-fidelity, expressive zero-shot voice cloning. Achieves near-human naturalness across multiple languages, outperforming proprietary models on standard TTS evaluations. 📄 Paper: https://arxiv.org/abs/2601.15621 📊 Benchmark: Seed-TTS Eval & MOS (WER, SIM-O, naturalness) - https://arxiv.org/abs/2601.15621 🔗 GitHub: https://github.com/QwenLM/Qwen3-TTS

microsoft/VibeVoice-1.5B

Text-to-Speech • 3B • Updated Jan 22 • 98.2k • 2.25k

Note 🏆 SOTA frontier open-source TTS by Microsoft (1.5B). Designed for expressive, long-form multi-speaker podcast generation with natural turn-taking and speaker consistency. Achieves human-level naturalness on TTS-Arena. Supports English and Chinese. 📄 Paper: https://arxiv.org/abs/2508.19205 📊 Benchmark: TTS-Arena & EmergentTTS-Eval (MOS, naturalness) - https://arxiv.org/abs/2508.19205 🔗 GitHub: https://github.com/microsoft/VibeVoice

fishaudio/s2-pro

Text-to-Speech • 5B • Updated 15 days ago • 15.3k • 747

Note 🏆 SOTA open-source TTS by Fish Audio (S2, 5B). Features multi-speaker, multi-turn generation and instruction-following via natural language. Achieves RTF of 0.195 with time-to-first-audio below 100ms — production-ready streaming inference. 📄 Paper: https://arxiv.org/abs/2603.08823 📊 Benchmark: EmergentTTS-Eval & Seed-TTS Eval (WER, SIM-O, expressiveness) - https://arxiv.org/abs/2603.08823 🔗 GitHub: https://github.com/fishaudio/fish-speech

IndexTeam/IndexTTS-2

Text-to-Speech • Updated Jan 20 • 17.9k • 660

Note 🏆 SOTA emotionally expressive auto-regressive zero-shot TTS (IndexTeam). Breakthrough in duration control and fine-grained emotion transfer. Achieves top scores on naturalness and speaker similarity benchmarks vs. proprietary models. 📄 Paper: https://arxiv.org/abs/2506.21619 📊 Benchmark: Seed-TTS Eval & MOS (WER, SIM-O, emotion accuracy) - https://arxiv.org/abs/2506.21619 🔗 GitHub: https://github.com/index-tts/index-tts

bosonai/higgs-audio-v2-generation-3B-base

Text-to-Speech • 6B • Updated Jul 28, 2025 • 609k • 663

Note 🏆 SOTA expressive audio foundation model by Boson AI (6B). Pretrained on 10M+ hours of audio data with deep language and acoustic understanding. Achieves #1 win rate on EmergentTTS-Eval for expressive speech, outperforming GPT-4o Audio. 📄 Paper: https://arxiv.org/abs/2505.23009 📊 Benchmark: EmergentTTS-Eval (win rate vs. GPT-4o Audio) - https://arxiv.org/abs/2505.23009 🔗 GitHub: https://github.com/boson-ai/higgs-audio

FunAudioLLM/Fun-CosyVoice3-0.5B-2512

Text-to-Speech • Updated Feb 3 • 6.14k • 491

Note 🏆 SOTA scalable multilingual zero-shot TTS by Alibaba FunAudio (CosyVoice3, 0.5B). LLM-based with supervised semantic tokens enabling natural prosody, cross-lingual voice cloning, and fine-grained style control across multiple languages. 📄 Paper: https://arxiv.org/abs/2407.05407 📊 Benchmark: Seed-TTS Eval & multilingual MOS (WER, SIM-O, DNSMOS) - https://arxiv.org/abs/2407.05407 🔗 GitHub: https://github.com/FunAudioLLM/CosyVoice

HKUSTAudio/Llasa-3B

Text-to-Speech • 4B • Updated May 10, 2025 • 486 • 526

Note 🏆 SOTA Llama-based speech synthesis scaling (HKUST, 3B). First TTS model to systematically scale both train-time and inference-time compute. Achieves SOTA on Seed-TTS Eval English & Chinese benchmarks, surpassing prior open-source models. 📄 Paper: https://arxiv.org/abs/2502.04128 📊 Benchmark: Seed-TTS Eval (WER, SIM-O) English & Chinese - https://arxiv.org/abs/2502.04128 🔗 GitHub: https://github.com/HKUSTAudio/Llasa

myshell-ai/OpenVoiceV2

Text-to-Speech • Updated Dec 24, 2024 • 478

Note 🏆 SOTA zero-shot cross-lingual voice cloning by MyShell AI (OpenVoice V2). Natively supports English, Spanish, French, Chinese, Japanese & Korean with accurate tone color cloning, flexible style control (emotion, rhythm, intonation). MIT license. 📄 Paper: https://arxiv.org/abs/2312.01479 📊 Benchmark: Speaker similarity & naturalness MOS vs. XTTS - https://arxiv.org/abs/2312.01479 🔗 GitHub: https://github.com/myshell-ai/OpenVoice

amphion/MaskGCT

Text-to-Speech • Updated Apr 13, 2025 • 731 • 306

Note 🏆 SOTA zero-shot TTS with Masked Generative Codec Transformer (Amphion). Non-autoregressive approach with parallel decoding for fast, high-quality synthesis. Achieves top scores on TTS-Arena for naturalness across 6 languages (EN, ZH, FR, DE, JA, KO). 📄 Paper: https://arxiv.org/abs/2409.00750 📊 Benchmark: TTS-Arena & SeedTTS Eval (MOS, SIM-O, WER) - https://arxiv.org/abs/2409.00750 🔗 GitHub: https://github.com/open-mmlab/Amphion