Writing study: VibeVoice: Expressive Podcast Generation with Next-Token Diffusion

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Zhiliang Peng, Jianwei Yu, Wenhui Wang, Yaoyao Chang, Yutao Sun, Li Dong, Yi Zhu, Weijiang Xu, Hangbo Bao, Zehua Wang, Shaohan Huang, Yan Xia, Furu Wei · Official ICLR 2026 oral record

Observed writing choices, not acceptance rules. This study examines the inspected manuscript version. It does not establish why the paper was accepted or that its presentation choices should be copied unchanged.

Figure, table and section ledger

Editorial contribution type: Architecture & representation design (primary).

Layout extraction, not a required paper template. Counts are distinct caption labels; a multi-panel figure counts once. Section spans are estimated from heading positions, including floats and whitespace. Method and limitations.

Canonical PDF

22 physical pages

Before references / appendix

≈ 9.16 page-equivalents

Main figures / tables

3 / 3

Attached appendix figures / tables

3 / 8

Attached appendix span

≈ 8.67 page-equivalents

Before references / appendix: ≈ 41.64%References / intervening material: ≈ 18.95%Attached appendix: ≈ 39.41%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

Section-by-section lengths and source pages

The title/abstract/front matter precede the first numbered heading. A span ends at the next detected top-level heading, not at the last sentence; end-page whitespace is included. References and statements are listed separately. These are not measurements of text-only or artwork area.

Top-level heading	Starts on PDF page	Estimated span	Share of entire PDF
1 INTRODUCTION	1	≈ 1.23 pp.	5.59%
2 METHOD	2	≈ 2.46 pp.	11.18%
3 EXPERIMENT	5	≈ 1.48 pp.	6.73%
4 RESULTS	6	≈ 2.12 pp.	9.64%
5 RELATED WORK	8	≈ 0.92 pp.	4.18%
6 CONCLUSION	9	≈ 0.38 pp.	1.73%
REFERENCES	10	≈ 4.17 pp.	18.95%
A DATA PREPARATION PIPELINE	14	≈ 0.43 pp.	1.95%
B MORE TOKENIZER ABLATIONS	14	≈ 1.49 pp.	6.77%
C MORE BENCHMARK RESULTS	16	≈ 0.35 pp.	1.59%
D MORE RECONSTRUCTION RESULTS	16	≈ 0.26 pp.	1.18%
E VIBEVOICE INFERENCE TIME	16	≈ 1.29 pp.	5.86%
F TRAINING HYPER-PARAMETERS	18	≈ 0.61 pp.	2.77%
G VIBEVOICE-OBJECTIVE EVALUATION	18	≈ 0.99 pp.	4.5%
H SUBJECTIVE EVALUATION	19	≈ 1.71 pp.	7.77%
I FIRST-PHASE EVALUATION RESULTS	21	≈ 0.75 pp.	3.41%
J SPECTRAL ANALYSIS OF DIFFUSION INFERENCE STEPS	22	≈ 0.48 pp.	2.18%
LLM USAGE STATEMENT	22	≈ 0.3 pp.	1.36%

All detected figure and table locations

Main text

• Figure 1 · p. 2
• Figure 2 · p. 3
• Table 1 · p. 6
• Table 2 · p. 7
• Figure 3 · p. 8
• Table 3 · p. 8

Attached appendix

• Table 4 · p. 15
• Figure 4 · p. 15
• Table 5 · p. 16
• Table 6 · p. 16
• Table 7 · p. 17
• Table 8 · p. 17
• Table 9 · p. 18
• Table 10 · p. 19
• Figure 5 · p. 20
• Table 11 · p. 21
• Figure 6 · p. 22

Argument and organization

1. From speech clips to conversations

   Motivates low-rate representations and persistent speaker conditioning.

2. Continuous tokenization and next-token diffusion

   Separates waveform fidelity, semantic feedback and language-model conditioning.

3. Long-form evaluation and component checks

   Tests conversation generation separately from short speech and codec reconstruction.

Problem, claim and contribution

The introduction motivates full conversations rather than isolated utterances, connecting duration to token rate and speaker consistency. The 90-minute capability claim should remain separate from the systematic benchmark, which spans one to thirty minutes.

§1, p.1; §4.1, pp.5–6; AppendixG, p.20 · Main paper

Argument and section order

Tokenizer design precedes the language model, diffusion head and curriculum, so each later module has a defined input representation. Continuous 7.5-Hz latents must not be described as discrete codes or as a bitrate directly comparable to multi-codebook codecs.

§§3.1–3.3, pp.3–5; Table3, p.8 · Main paper

Figures, captions and visual explanation

Figure1 distinguishes acoustic and semantic streams with persistent token colors and shows semantic feedback obtained by decoding and re-encoding generated audio. This is an unusually useful diagram of the actual autoregressive loop, not merely a component inventory.

Figure1 and §§3.1–3.2, pp.2–4 · Main paper

Tables, comparisons and evidence

The human study uses eight two-speaker scripts and twenty-four listeners, whereas objective tests cover additional speaker counts. Seven-billion parameters improves rated quality but does not improve every WER comparison; the reported rating spread does not establish statistical significance.

Table1 and §4.2, pp.6–7; AppendixH, pp.20–21 · Main paper

Appendix and reproducibility

The appendix documents an unsuccessful coupled-tokenizer alternative and decomposes measured latency. It also clarifies that streaming audio still conditions on the full script, and that one-step timing is not the default ten-step quality setting.

AppendicesB andE, pp.14–18; Tables6–7 · Main paper

Clarity and writing conventions

Figure3 separates fidelity and speaker similarity across sampling settings instead of implying more steps always help. Mooncast comparisons retain only successful generations for some multi-speaker cases, an important denominator restriction that must accompany favorable long-form results.

Figure3, p.8; §4.3 and Table2, pp.7–8 · Main paper

What to reuse—and what not to infer

Transferable lesson: Use a token-colored rollout diagram to explain feedback, then separate capability demonstrations, controlled benchmarks, human ratings and measured runtime rather than combining them into one quality claim.

Boundary: The internal podcast training corpus, limited human scripts, baseline exclusions and unmatched codec bitrates restrict broad comparisons. Qualitative denoising explanations are hypotheses, not causal evidence; voice-cloning capability also requires consent-aware deployment.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-09-19

Reading status

Full text + selected visuals inspected

Main text

pp.1–10, §§1–6; references pp.10–14

Appendix / supplement

pp.14–22, A–J and LLM-use disclosure: coupled-tokenizer alternative, short speech, codecs, timing, training, data, human evaluation and qualitative denoising

Visually inspected PDF pages

2, 8

• official ICLR 2026 proceedings PDF · 22 pages