Research map / Image generation / Autoregressive and latent generation
StyleTTS 2: Towards Human-Level Text-to-Speech through Style Diffusion and Adversarial Training with Large Speech Language Models
Yinghao Aaron Li, Cong Han, Vinay S Raghavan, Gavin Mischler, Nima Mesgarani
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add style diffusion (vs random encoded styles) (ablation) | StyleTTS 2 · LJSpeech OOD texts | CMOS-N: -0.46 | Improved |
| add SLM (WavLM) adversarial training (ablation) | StyleTTS 2 · LJSpeech OOD texts | CMOS-N: -0.32 | Improved |
| add prosodic style encoder (ablation) | StyleTTS 2 · LJSpeech OOD texts | CMOS-N: -0.35 | Improved |
| add differentiable duration upsampler (ablation) | StyleTTS 2 · LJSpeech OOD texts | CMOS-N: -0.21 | Improved |
| add OOD texts to SLM adversarial training (ablation) | StyleTTS 2 · LJSpeech OOD texts | CMOS-N: -0.15 | Improved |
| add SLM adversarial training (ablation) | StyleTTS 2 · LJSpeech | WER: WER 6.52 -> 6.50 | No clear effect |
| add SLM adversarial training (ablation) | StyleTTS 2 · LJSpeech | DUR MAD: DUR MAD 0.513 (w/o) vs 0.521 | Hurt |
| add differentiable upsampler (ablation) | StyleTTS 2 · LJSpeech | F0 RMSE: F0 RMSE 0.880 -> 0.651 | Improved |
| add style diffusion (ablation) | StyleTTS 2 · LJSpeech | WER: WER 8.77% -> 6.50% | Improved |
| increase style diffusion steps beyond 3-5 (ablation) | StyleTTS 2 · LJSpeech | MCD | No clear effect |
| use StyleTTS 2 instead of ground truth recordings | TTS system · LJSpeech | CMOS-N: 0.28 | Improved |
| use StyleTTS 2 vs ground truth | TTS system · VCTK | CMOS-N: -0.02 | No clear effect |
Later papers that compared with this method
2 later accepted papers report StyleTTS2 as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- MambaVoiceCloning: Efficient and Expressive Text-to-Speech via State-Space Modeling and Diffusion Control (ICLR 2026 · better on 9 of 11 shared results)
- NaturalSpeech 3: Zero-Shot Speech Synthesis with Factorized Codec and Diffusion Models (ICML 2024 · better on 1 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- NaturalSpeech 2: Latent Diffusion Models are Natural and Zero-Shot Speech and Singing Synthesizers (ICLR 2024 · similarity 0.88)
- DELTA-TTS: Adapting Autoregressive Model into a Diffusion Language Model for Text-to-Speech (NeurIPS 2026 · similarity 0.87)
- DiTTo-TTS: Diffusion Transformers for Scalable Text-to-Speech without Domain-Specific Factors (ICLR 2025 · similarity 0.87)
- MM-TTS: Multi-Modal Prompt Based Style Transfer for Expressive Text-to-Speech Synthesis (AAAI 2024 · similarity 0.86)
- StyleSinger: Style Transfer for Out-of-Domain Singing Voice Synthesis (AAAI 2024 · similarity 0.86)
- DDDM-VC: Decoupled Denoising Diffusion Models with Disentangled Representation and Prior Mixup for Verified Robust Voice Conversion (AAAI 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
