Research map / Image generation / Autoregressive and latent generation
Diffusion audio synthesis
Generate waveforms and general audio with diffusion and latent diffusion models.
52 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2017–2026) · an approach in Autoregressive and latent generation, keyword cluster “speech · music · audio”
State History Tried Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
17 papers in 2023–24 and 23 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 3.6% to 2.6%.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2017 · SampleRNN: An Unconditional End-to-End Neural Audio Generation Model (ICLR 2017 · 423 citations)
- 2018 · SING: Symbol-to-Instrument Neural Generator (NeurIPS 2018 · 25 citations)
- 2019 · ClariNet: Parallel Wave Generation in End-to-End Text-to-Speech (ICLR 2019 · 62 citations)
- 2020 · DDSP: Differentiable Digital Signal Processing (ICLR 2020 · 75 citations)
- 2021 · DiffWave: A Versatile Diffusion Model for Audio Synthesis (ICLR 2021 · 119 citations)
- 2022 · DiffSinger: Singing Voice Synthesis via Shallow Diffusion Mechanism (AAAI 2022 · 197 citations)
- 2023 · DiffTalk: Crafting Diffusion Models for Generalized Audio-Driven Portraits Animation (CVPR 2023 · 128 citations)
- 2024 · Seeing and Hearing: Open-domain Visual-Audio Generation with Diffusion Latent Aligners (CVPR 2024 · 37 citations)
- 2025 · Enhancing Dance-to-Music Generation via Negative Conditioning Latent Diffusion Model (CVPR 2025 · 4 citations)
- 2026 · PASE: Leveraging the Phonological Prior of WavLM for Low-Hallucination Generative Speech Enhancement (AAAI 2026 · 3 citations)
Earliest papers here
- SampleRNN: An Unconditional End-to-End Neural Audio Generation Model (ICLR 2017 · 423 citations)
- Neural Audio Synthesis of Musical Notes with WaveNet Autoencoders (ICML 2017 · 390 citations)
- SING: Symbol-to-Instrument Neural Generator (NeurIPS 2018 · 25 citations)
What has been tried
36 papers here report 433 measured changes (a component added, removed or swapped, compared with a baseline): helped 68%, no clear effect 8%, hurt 21%, mixed 3%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| add implicit phase wrapping instead of predicting absolute phase with scaled tanh | LibriTTS · UTMOS | 3.590 -> 3.734 | Helped | Vocos: Closing the gap between time-domain and Fourier-based neural vocoders for high-quality audio synthesis (ICLR 2024 · 11 citations) |
| Normalize the PIA initialization (PIAN) instead of using unnormalized PIA | LJSpeech, VCTK, and LibriTTS · TPR@1% FPR | 94.2 -> 95.7; 20.6 -> 19.6; 30.0 -> 44.7 | Mixed | An Efficient Membership Inference Attack for the Diffusion Model by Proximal Initialization (ICLR 2024 · 9 citations) |
| add correction step | Slakh2100 test set, all stems · SI-SDR_I (dB) | 16.30 -> 17.27 | Helped | Multi-Source Diffusion Models for Simultaneous Music Generation and Separation (ICLR 2024 · 8 citations) |
| Use CLAP_ours instead of CLAP_LAION text embeddings | MusicCaps · FD_openl3 | 123.30 -> 118.09 | Helped | Fast Timing-Conditioned Latent Audio Diffusion (ICML 2024 · 7 citations) |
| Combine visual rhythm and motion conditioning rather than motion conditioning alone | AIST++ · F1 | 91.59 -> 98.46 | Helped | Enhancing Dance-to-Music Generation via Negative Conditioning Latent Diffusion Model (CVPR 2025 · 4 citations) |
| Add restricted temporal context for codebooks above the first | In-domain test set, span length 3 · FAD | 0.8221 -> 0.61 | Helped | Masked Audio Generation using a Single Non-Autoregressive Transformer (ICLR 2024 · 2 citations) |
| Use track-specific adaptive timestep conditioning instead of canonical single-sample inpainting | Slakh2100 S_A test, drums extraction · Log-Mel L1 distance | 3.20 -> 2.71 | Helped | MGE-LDM: Joint Latent Diffusion for Simultaneous Music Generation and Source Extraction (NeurIPS 2025 · 1 citation) |
| Add energy conditioning | LJ-Speech test with true phonemes and durations · CER (%) | 2.90 -> 1.18 | Helped | DiffAR: Denoising Diffusion Autoregressive Model for Raw Speech Waveform Generation (ICLR 2024 · 1 citation) |
| Use historical context patches in LocDiT instead of removing them | Seed-EN, patch size 4 · WER (%) | 22.874^\star -> \textbf{1.736} | Helped | DiTAR: Diffusion Transformer Autoregressive Modeling for Speech Generation (ICML 2025 · 0 citations) |
| Add outpainting and inpainting | Ablation study · FAD | 34.17 -> 8.00 | Helped | BNMusic: Blending Environmental Noises into Personalized Music (NeurIPS 2025 · 0 citations) |
| replace the target reverse prompt with a random prompt | MUSIC test · C-A | 0.615 -> 0.363 | Hurt | ZeroSep: Separate Anything in Audio with Zero Training (NeurIPS 2025 · 0 citations) |
| Use clustering to group tokens before watermarking | MMW Book Report · Mismatch rate | 0.3749 -> 0.2117 | Helped | Robust Distortion-Free Watermark for Autoregressive Audio Generation Models (NeurIPS 2025 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 14 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- MUSDB18 · 3 of 14 papers
- ESC-50 · 3 of 14 papers
- FSD50K · 3 of 14 papers
- AudioCaps · 3 of 14 papers
- Emilia · 2 of 14 papers
- WavCaps · 2 of 14 papers
- VGGSound · 2 of 14 papers
- MTG-Jamendo · 2 of 14 papers
Baselines
- Stable Audio Open · 3 of 14 papers
- F5TTS · 2 of 14 papers
- AudioSep · 2 of 14 papers
- pitch shifting · 2 of 14 papers
- UniAudio · 2 of 14 papers
- MusicGen-Large · 2 of 14 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 40 papers; from OpenAlex and the papers' affiliations).
- Meta · 3 papers
- Microsoft · 2 papers
- University of Montreal · 2 papers
- Tencent · 2 papers
Use it
Download BibTeX of 37 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Neural text-to-speech synthesis (72 papers)
- Controllable music generation (42 papers)
Elsewhere in the topic
- Classical generative image synthesis (147 papers)
- High-resolution diffusion synthesis (140 papers)
- Retrieval and matching with generative features (140 papers)
- Generative latent space tooling (129 papers)
- Generalizable fake image detection (119 papers)
- Latent diffusion with learned tokenizers (114 papers)
- Synthetic training data from generators (111 papers)
- Compositional scene generation (105 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).