Research map / Label-efficient and robust learning / Self-supervised and contrastive
Self-supervised Neural Factor Analysis for Disentangling Utterance-level Speech Representations
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 |
|---|---|---|---|
| gradient-based ELBO training of the factor analysis model (vs. EM-trained loading matrix) (ablation) | HuBERT Large features · VoxCeleb1 ASV | EER: 2.54% -> 2.26% | Helped |
| gradient-based ELBO training instead of EM (ablation) | HuBERT Large features · IEMOCAP ER | accuracy: 73.4% -> 78.1% | Helped |
| gradient-based ELBO training instead of EM (ablation) | XLS-R features · Common Language LID | accuracy: 83.6% -> 86.3% | Helped |
| NFA with gradient-based ELBO (HuBERT-based, large) vs WavLM Large | HuBERT Large · VoxCeleb1 SID / SD / ASV | EER/DER/accuracy: SID 95.49 -> 98.1; SD 3.24 -> 1.84; ASV 3.77 -> 2.26 | Helped |
| NFA ELBO training on HuBERT base features | HuBERT base · Zero-shot LibriSpeech SV | EER: 28.7 -> 3.98 | Helped |
| NFA ELBO training on HuBERT base features | HuBERT base · Zero-shot VoxCeleb1 SV | EER: 32.1 -> 9.32 | Helped |
| NFA ELBO training on HuBERT features (ASR preservation check) | HuBERT base · LibriSpeech clean ASR | WER: 6.42 -> 6.31 | No clear effect |
| NFA ELBO training on HuBERT Large features (ASR preservation check) | HuBERT Large · LibriSpeech clean ASR | WER: 3.62 -> 3.66 | No clear effect |
| NFA ELBO training with 10-20% labeled data | HuBERT Large / XLS-R NFA · Low label-resource (20% labels) ASV, SD, SID, ER, LID | task performance | Helped |
| Ablation of the utterance-level ELBO-trained objective isn't reported; mean-averaging of frame features (no factor analysis) for zero-shot SV | HuBERT base · Zero-shot LibriSpeech SV | EER: HuBERT 28.7 vs I-vector 11.2 | Hurt |
| Domain shift: NFA trained on LibriSpeech evaluated on out-of-domain VoxCeleb1/VOiCES | NFA base · Zero-shot VoxCeleb1 / VOiCES | EER: 3.98 (LibriSpeech) -> 9.32 (VoxCeleb1) -> 12.32 (VOiCES) | Hurt |
Closest papers by meaning
- Self-Supervised Disentangled Representation Learning for Robust Target Speech Extraction (AAAI 2024 · similarity 0.88)
- CA-SSLR: Condition-Aware Self-Supervised Learning Representation for Generalized Speech Processing (NeurIPS 2024 · similarity 0.87)
- Self-Taught Recognizer: Toward Unsupervised Adaptation for Speech Foundation Models (NeurIPS 2024 · similarity 0.86)
- SelfVC: Voice Conversion With Iterative Refinement using Self Transformations (ICML 2024 · similarity 0.86)
- Unified Speech Recognition: A Single Model for Auditory, Visual, and Audiovisual Inputs (NeurIPS 2024 · similarity 0.85)
- Speech Self-Supervised Learning Using Diffusion Model Synthetic Data (ICML 2024 · similarity 0.85)
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.