Research map / Probabilistic and scientific modeling / Variational and generative sampling
Re-examination of the Role of Latent Variables in Sequence Modeling
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Replace a fully factorized output with an autoregressive output | RNN · Speech datasets (TIMIT, VCTK, Blizzard) | log-likelihood: 32,745 -> 117,721; 0.786 -> 3.2173; 7,610 -> 22,714 | Helped |
| Replace a fully factorized output with an autoregressive output | RNN · Muse and Nottingham | log-likelihood: -6.991 -> -5.161; -3.400 -> -2.028 | Helped |
| Replace a fully factorized output with an autoregressive output | RNN · Perm-TIMIT | log-likelihood: 25,679 -> 15,763 | Hurt |
| Replace a fully factorized output with an autoregressive output | RNN · IAM-OnDB | log-likelihood: 1397 -> 1440 | Helped |
| Use deterministic RNNs instead of SRNNs with autoregressive output distributions | Autoregressive RNN and SRNN models · TIMIT, VCTK, Blizzard, Muse, Nottingham, IAM-OnDB, and Perm-TIMIT | log-likelihood: RNN-hier vs. SRNN-hier: 109,641 vs. 107,912; 3.1822 vs. 3.1423; 21,950 vs. 21,845; -5.161 vs. -5.483; -2.028 vs. -2.065; | Helped |
| Use the flat architecture instead of the hierarchical architecture | RNN · TIMIT, VCTK, and Blizzard | log-likelihood: 109,641 -> 117,721; 3.1822 -> 3.2173; 21,950 -> 22,714 | Helped |
| Use the hierarchical architecture instead of the flat architecture | RNN · Muse, Nottingham, and Perm-TIMIT | log-likelihood: -5.251 -> -5.161; -2.180 -> -2.028; 15,763 -> 95,161 | Helped |
| Replace RNN with SRNN on stride-50 and stride-200 scalar-frame sequences | RNN · Stride-TIMIT, stride-VCTK, and stride-Blizzard | log-likelihood: Stride 50: 20,655 -> 14,469; 0.668 -> 0.605; 4,607 -> 3,603. Stride 200: 4,124 -> -1,137; 0.177 -> 0.0187; -320 -> -1,23 | Hurt |
| Use interleaved delta copying with U=2 instead of a factorized RNN | F-RNN · TIMIT, VCTK, and Blizzard | log-likelihood: 32,745 -> 70,900; 0.786 -> 2.027; 7,610 -> 15,306 | Helped |
| Use interleaved delta copying with U=3 instead of a factorized RNN | F-RNN · TIMIT, VCTK, and Blizzard | log-likelihood: 32,745 -> 72,067; 0.786 -> 2.262; 7,610 -> 15,284 | Helped |
| Use random delta copying with V=50 instead of a factorized RNN | F-RNN · TIMIT, VCTK, and Blizzard | log-likelihood: 32,745 -> 66,122; 0.786 -> 2.199; 7,610 -> 14,389 | Helped |
| Use random delta copying with V=75 instead of a factorized RNN | F-RNN · TIMIT, VCTK, and Blizzard | log-likelihood: 32,745 -> 66,453; 0.786 -> 2.120; 7,610 -> 14,585 | Helped |
Closest papers by meaning
- Z-Forcing: Training Stochastic Recurrent Networks (NeurIPS 2017 · similarity 0.90)
- Continuous Autoregressive Modeling with Stochastic Monotonic Alignment for Speech Synthesis (ICLR 2025 · similarity 0.86)
- NaturalSpeech 2: Latent Diffusion Models are Natural and Zero-Shot Speech and Singing Synthesizers (ICLR 2024 · similarity 0.84)
- DiffAR: Denoising Diffusion Autoregressive Model for Raw Speech Waveform Generation (ICLR 2024 · similarity 0.84)
- Dynamic-SUPERB Phase-2: A Collaboratively Expanding Benchmark for Measuring the Capabilities of Spoken Language Models with 180 Tasks (ICLR 2025 · similarity 0.84)
- E-BATS: Efficient Backpropagation-Free Test-Time Adaptation for Speech Foundation Models (NeurIPS 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.