Research map / Image generation / Autoregressive and latent generation
KALL-E: Autoregressive Speech Synthesis with Next-Distribution Prediction
Kangxiang Xia, Xinfa Zhu, Jixun Yao, Wenjie Tian, Wenhao Li, Lei Xie
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 test-time training on one reference utterance | KALL-E · Seed-TTS test-zh | SIM: 0.646 -> 0.698 | Improved |
| add test-time training | KALL-E · Seed-TTS test-en | SIM: 0.568 -> 0.611 | Improved |
| add test-time training | KALL-E · Seed-TTS test-zh | CER: 0.96 -> 1.02 | No clear effect |
| use Flow-VAE instead of Stable Audio VAE at 64-dim 12.5Hz for the LM (ablation) | KALL-E LM · small-scale TTS | CER: 40.09 -> 6.15 | Improved |
| increase Flow-VAE latent dim from 64 to 512 (ablation) | KALL-E LM · small-scale TTS | CER: 6.15 -> 2.79 | Improved |
| increase Flow-VAE latent dim from 512 to 1024 (ablation) | KALL-E LM · small-scale TTS | CER: 2.79 -> 19.69 | Hurt |
| raise frame rate from 12.5 to 25 Hz at 256-dim (ablation) | KALL-E LM · small-scale TTS | CER: 4.91 -> 7.30 | Hurt |
| increase TTT set size N beyond about 200 (ablation) | KALL-E (TTT) · 200-item Seed-TTS zh subset | CER | Hurt |
| use 12.5 Hz frame rate vs baselines at 25-93.75 Hz | KALL-E · 10s audio inference | GFLOPs: 7 947.48 vs 122 170.02 (Llasa 1B) | Improved |
| lower Flow-VAE latent dim from 1024 to 64 at 12.5 Hz (ablation) | Flow-VAE · LibriSpeech test-clean | PESQ WB: 3.71 -> 2.01 | Hurt |
Closest papers by meaning
- DELTA-TTS: Adapting Autoregressive Model into a Diffusion Language Model for Text-to-Speech (NeurIPS 2026 · similarity 0.91)
- Continuous Autoregressive Modeling with Stochastic Monotonic Alignment for Speech Synthesis (ICLR 2025 · similarity 0.90)
- DiTAR: Diffusion Transformer Autoregressive Modeling for Speech Generation (ICML 2025 · similarity 0.88)
- Efficient Speech Language Modeling via Energy Distance in Continuous Latent Space (NeurIPS 2025 · similarity 0.87)
- HALL-E: Hierarchical Neural Codec Language Model for Minute-Long Zero-Shot Text-to-Speech Synthesis (ICLR 2025 · similarity 0.87)
- Generative Audio Language Modeling with Continuous-valued Tokens and Masked Next-Token Prediction (ICML 2025 · 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.
