Research map / Probabilistic and scientific modeling / Variational and generative sampling
Unsupervised Learning of Disentangled and Interpretable Representations from Sequential Data
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 |
|---|---|---|---|
| use sequence-level μ₂ rather than segment-based μ₁ for speaker representation (ablation) | FHVAE · TIMIT speaker verification | EER (Raw): 22.77% -> 3.87% | Helped |
| increase discriminative weight α (ablation) | FHVAE μ₂, dimension 16 · TIMIT speaker verification | EER (Raw): 5.06% -> 2.38% | Helped |
| use FHVAE z₁ features instead of FBank | ASR trained on male speakers · TIMIT female test speakers | PER: 32.8% -> 26.2% | Helped |
| use FHVAE z₁ features instead of FBank | ASR trained on clean speech · Aurora-4 noisy test | WER: 50.97% -> 16.42% | Helped |
| use FHVAE z₁ features instead of FBank | ASR trained on clean speech · Aurora-4 channel test | WER: 36.99% -> 20.29% | Helped |
| use FHVAE z₁ features instead of FBank | ASR trained on clean speech · Aurora-4 NC test | WER: 71.80% -> 36.33% | Helped |
| use latent sequence z₂ instead of latent segment z₁ as ASR features (ablation) | FHVAE ASR · Aurora-4 NC test | WER: 36.33% -> 86.36% | Hurt |
| exclude the NC domain from FHVAE training (ablation) | FHVAE z₁ ASR features · Aurora-4 NC test | WER: 36.33% -> 40.59% | Hurt |
| replace LSTM with a recurrent encoder/decoder rather than a feed-forward network (ablation) | Seq2Seq-FHVAE · TIMIT test | segment variational lower bound: -348.63 -> -143.80 | Helped |
| replace LSTM with GRU (ablation) | Seq2Seq-FHVAE · TIMIT test | segment variational lower bound: -143.80 -> -158.42 | Hurt |
| use FHVAE z₁ instead of FBank | ASR trained on clean speech · Aurora-4 clean test | WER: 3.47% -> 5.01% | Hurt |
Closest papers by meaning
- Disentangled Sequential Autoencoder (ICML 2018 · similarity 0.88)
- Sequential Disentanglement by Extracting Static Information From A Single Sequence Element (ICML 2024 · similarity 0.88)
- Self-supervised Neural Factor Analysis for Disentangling Utterance-level Speech Representations (ICML 2023 · similarity 0.88)
- SASA: Subspace-Aware Sparse Autoencoders for Effective Mechanistic Interpretability (NeurIPS 2026 · similarity 0.86)
- Incorporating Hierarchical Semantics in Sparse Autoencoder Architectures (NeurIPS 2026 · similarity 0.86)
- Self-Supervised Disentangled Representation Learning for Robust Target Speech Extraction (AAAI 2024 · 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.