In this Github repository, we summarize popular open-sourced TTS models from the past three years, which primarily fall into two categories: autoregressive (AR) and non-autoregressive (NAR) architectures. Note that we do not cover TTS in end-to-end speech LLMs.
Text-to-Speech (TTS) is a transformative technology that converts written text into natural-sounding, human-like speech. The core objective of a TTS system is to produce synthetic speech that is not only intelligible but also fluent, expressive, and indistinguishable from a real human voice.
Based on the type of decoder, AR TTS models can be divided into two categories: flow/diffusion decoder and codec decoder.
| Model | Title | Tokenizer | Paper | Github |
|---|---|---|---|---|
| CosyVoice series | CosyVoice 3: Towards In-the-wild Speech Generation via Scaling-up and Post-training | single-codebook discrete token 25Hz | Link1 Link2 Link3 | Link |
| GLM-TTS | GLM-TTS Technical Report | single-codebook discrete token 25Hz | Link | Link |
| IndexTTS2 | IndexTTS2: A Breakthrough in Emotionally Expressive and Duration-Controlled Auto-Regressive Zero-Shot Text-to-Speech | single-codebook discrete token 25Hz | Link | Link |
| VoxCPM | VoxCPM: Tokenizer-Free TTS for Context-Aware Speech Generation and True-to-Life Voice Cloning | continuous features 12.5/6.25Hz | Link | Link |
| VibeVoice | VibeVoice Technical Report | continuous features 7.5Hz | Link | Link |
| Ming-omni-tts | Ming-Omni: A Unified Multimodal Model for Perception and Generation | continuous features 12.5Hz (Patch 3.1Hz) | Link | Link |
| Model | Title | Tokenizer | Paper | Github |
|---|---|---|---|---|
| Qwen3-TTS | Qwen3-TTS Technical Report | multi-codebook discrete token 25Hz | Link | Link |
| FireRedTTS2 | FireRedTTS-2: Towards Long Conversational Speech Generation for Podcast and Chatbot | multi-codebook discrete token 12.5Hz | Link | Link |
| Fish Audio S2 | Fish Audio S2 Technical Report | multi-codebook discrete token 21Hz | Link | Link |
| HiggsAudio | Higgs Audio: Redefining Expressiveness in Audio Generation | multi-codebook discrete token 25Hz | Link | Link |
| MOSS-TTS | MOSS-TTS Technical Report | multi-codebook discrete token 12.5Hz | Link | Link |
| Spark-TTS | Spark-TTS: An Efficient LLM-Based Text-to-Speech Model with Single-Stream Decoupled Speech Tokens | single-codebook discrete token 50Hz | Link | Link |
Based on the type of decoder, NAR TTS models can be also divided into three categories: traditional decoder (vocoder), codec decoder and VAE decoder.
| Model | Title | Paper | Github |
|---|---|---|---|
| F5-TTS | F5-TTS: A Fairytaler that Fakes Fluent and Faithful Speech with Flow Matching | Link | Link |
| ZipVoice | ZipVoice: Fast and High-Quality Zero-Shot Text-to-Speech with Flow Matching | Link | Link |
| Model | Title | Tokenizer | Paper | Github |
|---|---|---|---|---|
| MaskGCT | MaskGCT: Zero-Shot Text-to-Speech with Masked Generative Codec Transformer | multi-codebook discrete token 50Hz | Link | Link |
| OmniVoice | OmniVoice: Towards Omnilingual Zero-Shot Text-to-Speech with Diffusion Language Models | multi-codebook discrete token 25Hz | Link | Link |
| Model | Title | Tokenizer | Paper | Github |
|---|---|---|---|---|
| MegaTTS3 | MegaTTS 3: Sparse Alignment Enhanced Latent Diffusion Transformer for Zero-Shot Speech Synthesis | continuous features 25Hz | Link | Link |
| LongCat-AudioDiT | LongCat-AudioDiT: High-Fidelity Diffusion Text-to-Speech in the Waveform Latent Space | continuous features 11.72Hz | Link | Link |