Research map / Image generation / Autoregressive and latent generation
Hierarchical Semantic-Acoustic Modeling via Semi-Discrete Residual Representations for Expressive End-to-End Speech Synthesis
Yixuan Zhou, Guoyang Zeng, Xin Liu, Xiang Li, Renjie Yu, Ziyang Wang, Runchuan Ye, Weiyue Sun, Jiancheng Gui, Kehan Li, Zhiyong Wu, Zhiyuan Liu
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 FSQ semi-discrete bottleneck (vs purely continuous) (ablation) | VoxCPM · ZH-hard (ablation) | CER (%): 24.92 -> 18.19 | Improved |
| add RALM residual acoustic model (ablation) | VoxCPM · EN (ablation) | WER (%): 4.34 -> 2.98 | Improved |
| increase single-stream TSLM capacity from 24 to 30 layers (no RALM) (ablation) | w/o RALM variant · EN (ablation) | WER (%): 4.34 -> 4.12 | No clear effect |
| add pre-trained text LM init to TSLM (ablation) | VoxCPM · EN (ablation) | WER (%): 5.24 -> 2.98 | Improved |
| add original acoustic embeddings input to RALM (ablation) | VoxCPM · EN (ablation) | WER (%): 4.91 -> 2.98 | Improved |
| add residual hidden state in LocDiT condition (ablation) | VoxCPM · EN (ablation) | WER (%): 3.86 -> 2.98 | Improved |
| FSQ dimension 4 instead of 256 (ablation) | VoxCPM · EN (ablation) | WER (%): 2.98 -> 5.18 | Hurt |
| FSQ dimension 128 instead of 256 (ablation) | VoxCPM · ZH-hard (ablation) | CER (%): 18.19 -> 16.76 | Mixed |
| replace sum fusion with attention fusion (ablation) | LocDiT conditioning · EN (ablation) | WER (%): 2.98 -> 4.92 | Hurt |
| replace sum fusion with concatenation (ablation) | LocDiT conditioning · ZH (ablation) | CER (%): 1.77 -> 2.88 | Mixed |
| remove classifier-free guidance (CFG 1.0 vs 2.0) (ablation) | VoxCPM · Seed-TTS-eval EN | WER (%): 1.85 -> 16.32 | Hurt |
| raise CFG from 2.0 to 5.0 (ablation) | VoxCPM · Seed-TTS-eval EN | WER (%): 1.85 -> 12.78 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- IndexTTS 2 — IndexTTS2: A Breakthrough in Emotionally Expressive and Duration-Controlled Auto-Regressive Zero-Shot Text-to-Speech (AAAI 2026) · better on 3 of 5 shared results
- MaskGCT — MaskGCT: Zero-Shot Text-to-Speech with Masked Generative Codec Transformer (ICLR 2025) · 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
- WavTokenizer: an Efficient Acoustic Discrete Codec Tokenizer for Audio Language Modeling (ICLR 2025 · similarity 0.88)
- TaDiCodec: Text-aware Diffusion Speech Tokenizer for Speech Language Modeling (NeurIPS 2025 · similarity 0.88)
- Generative Audio Language Modeling with Continuous-valued Tokens and Masked Next-Token Prediction (ICML 2025 · similarity 0.88)
- DELTA-TTS: Adapting Autoregressive Model into a Diffusion Language Model for Text-to-Speech (NeurIPS 2026 · similarity 0.88)
- AudioCALM: Continuous Autoregressive Language Modeling for Universal Audio Generation (NeurIPS 2026 · similarity 0.88)
- NaturalSpeech 2: Latent Diffusion Models are Natural and Zero-Shot Speech and Singing Synthesizers (ICLR 2024 · similarity 0.88)
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.
