Research map / Label-efficient and robust learning / Recognition and long-tail
Unsupervised Speech Recognition via Segmental Empirical Output Distribution Matching
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 |
|---|---|---|---|
| Refine phoneme boundaries using the current classifier in a second training iteration (ablation) | Segmental Empirical-ODM speech recognition · TIMIT fully unsupervised, matching LM | PER: 47.0 -> 42.6 | Helped |
| Refine phoneme boundaries using the current classifier in a second training iteration (ablation) | Segmental Empirical-ODM speech recognition · TIMIT fully unsupervised, non-matching LM | PER: 61.7 -> 49.1 | Helped |
| Add monophone HMM training (ablation) | Our Model: 2nd iteration · TIMIT fully unsupervised, matching LM | PER: 42.6 -> 41.5 | Helped |
| Use triphone instead of monophone HMM training (ablation) | Our Model: 2nd iteration · TIMIT fully unsupervised, matching LM | PER: 41.5 -> 39.4 | Helped |
| Add SAT to triphone HMM training (ablation) | Our Model: 2nd iteration · TIMIT fully unsupervised, matching LM | PER: 39.4 -> 36.5 | Helped |
| Add monophone HMM training (ablation) | Our Model: 2nd iteration · TIMIT fully unsupervised, non-matching LM | PER: 49.1 -> 44.7 | Helped |
| Use triphone instead of monophone HMM training (ablation) | Our Model: 2nd iteration · TIMIT fully unsupervised, non-matching LM | PER: 44.7 -> 44.9 | Hurt |
| Add SAT to triphone HMM training (ablation) | Our Model: 2nd iteration · TIMIT fully unsupervised, non-matching LM | PER: 44.9 -> 41.6 | Helped |
| Include the frame-wise smoothness term in Segmental Empirical-ODM (ablation) | Segmental Empirical-ODM · TIMIT fully unsupervised, matching LM | FER | Helped |
| Use a non-matching rather than matching phoneme language model | Our Model with oracle boundaries · TIMIT oracle boundaries | PER: 32.5 -> 40.1 | Hurt |
| Use a non-matching rather than matching phoneme language model | Our Model: 2nd iteration · TIMIT fully unsupervised | PER: 42.6 -> 49.1 | Hurt |
| Use the unsupervised method with oracle boundaries rather than fully unsupervised boundary estimation | Segmental Empirical-ODM speech recognition · TIMIT matching LM | PER: 42.6 -> 32.5 | Helped |
Closest papers by meaning
- REBORN: Reinforcement-Learned Boundary Segmentation with Iterative Training for Unsupervised ASR (NeurIPS 2024 · similarity 0.88)
- Unified Speech Recognition: A Single Model for Auditory, Visual, and Audiovisual Inputs (NeurIPS 2024 · similarity 0.85)
- Self-Taught Recognizer: Toward Unsupervised Adaptation for Speech Foundation Models (NeurIPS 2024 · similarity 0.85)
- SyllableLM: Learning Coarse Semantic Units for Speech Language Models (ICLR 2025 · similarity 0.84)
- UniWav: Towards Unified Pre-training for Speech Representation Learning and Generation (ICLR 2025 · similarity 0.83)
- Speech Recognition Meets Large Language Model: Benchmarking, Models, and Exploration (AAAI 2025 · similarity 0.83)
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.