Research map / Image generation / Autoregressive and latent generation
DMOSpeech 2: Reinforcement Learning for Duration Prediction in Metric-Optimized Speech Synthesis
Yinghao Aaron Li, Xilin Jiang, Fei Tao, Cheng Niu, Kaifeng Xu, Juntong Song, 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 GRPO reinforcement learning to the duration predictor (SIM and WER rewards) (ablation) | DMOSpeech 2 student (4 steps) · Seed-TTS test-en | WER: 3.750 -> 1.752 | Improved |
| add GRPO reinforcement learning to the duration predictor (ablation) | DMOSpeech 2 student (4 steps) · Seed-TTS test-en | SIM-o: 0.672 -> 0.698 | Improved |
| add GRPO reinforcement learning to the duration predictor (ablation) | DMOSpeech 2 student (4 steps) · Seed-TTS test-zh | CER: 2.000 -> 1.527 | Improved |
| add GRPO duration RL (subjective) (ablation) | DMOSpeech 2 student · Seed-TTS test-en CMOS | CMOS-N: -0.43 | Improved |
| replace 32-step F5-TTS teacher with 4-step distilled DMOSpeech 2 | F5-TTS · Seed-TTS RTF | RTF: 0.1671 -> 0.0316 | Improved |
| add teacher-guided sampling (14 teacher steps + 2 student steps) (ablation) | DMOSpeech 2 student · Pitch diversity CVf0 | CVf0: 0.4640 -> 0.5932 | Improved |
| add teacher-guided sampling (ablation) | DMOSpeech 2 student · Seed-TTS test-en | WER: 1.752 -> 1.738 | Improved |
| add teacher-guided sampling (ablation) | DMOSpeech 2 student · Seed-TTS RTF | RTF: 0.0316 -> 0.0941 | Hurt |
| add teacher-guided sampling (ablation) | DMOSpeech 2 student · Seed-TTS test-zh | CMOS-N: +0.45 (CMOS-N), +0.3 (CMOS-S) | Mixed |
| distill teacher into few-step student with distribution matching | F5-TTS teacher · Pitch diversity CVf0 | CVf0: 0.6659 -> 0.4640 | Hurt |
| use DMOSpeech 2 instead of 8B LLM-based TTS | LLaSA-8B comparison · Seed-TTS test-en | WER: 3.994 -> 1.752 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- MaskGCT — MaskGCT: Zero-Shot Text-to-Speech with Masked Generative Codec Transformer (ICLR 2025) · better on 0 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
- DMOSpeech: Direct Metric Optimization via Distilled Diffusion Model in Zero-Shot Speech Synthesis (ICML 2025 · similarity 0.91)
- Kinetic-Optimal Scheduling with Moment Correction for Metric-Induced Discrete Flow Matching in Zero-Shot Text-to-Speech (NeurIPS 2026 · similarity 0.84)
- NaturalSpeech 2: Latent Diffusion Models are Natural and Zero-Shot Speech and Singing Synthesizers (ICLR 2024 · similarity 0.84)
- TaDiCodec: Text-aware Diffusion Speech Tokenizer for Speech Language Modeling (NeurIPS 2025 · similarity 0.84)
- DiTTo-TTS: Diffusion Transformers for Scalable Text-to-Speech without Domain-Specific Factors (ICLR 2025 · similarity 0.84)
- DiTAR: Diffusion Transformer Autoregressive Modeling for Speech Generation (ICML 2025 · similarity 0.84)
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.
