Research map / Label-efficient and robust learning / Self-supervised and contrastive
Pay Attention to CTC: Fast and Robust Pseudo-Labelling for Unified Speech Recognition
Alexandros Haliassos, Rodrigo Mira, Stavros Petridis
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 USR autoregressive pseudo-labelling with CTC-driven teacher forcing plus mixed sampling (USR 2.0) | USR (Large, LRS3+Vox2) · LRS3 V (low-resource) | WER (%): 26.9 -> 23.7 | Improved |
| USR 2.0 vs USR at Base with LRS3 only | USR (Base) · LRS3 V (low-resource) | WER (%): 36.0 -> 36.2 | No clear effect |
| USR 2.0 vs USR under babble noise | USR (Base) · LRS3 ASR noisy, avg | WER (%): 43.3 -> 39.3 | Improved |
| add CTC PL supervision to the decoder in CTC-driven mode (ablation) | USR 2.0 · AVSpeech OOD AVSR | WER (%): 35.1 -> 24.2 | Improved |
| add attention-PL supervision to the decoder in CTC-driven mode (ablation) | USR 2.0 · LRS3 AVSR (ID) | WER (%): 3.6 -> 3.2 | Improved |
| supervise CTC head with attention PLs in CTC-driven mode (ablation) | USR 2.0 · LRS3 / AVSpeech AVSR | WER (%): 3.2/24.2 -> 3.2/24.3 | No clear effect |
| add attention PLs to CTC head in AR mode (ablation) | USR 2.0 · AVSpeech OOD AVSR | WER (%): 45.1 -> 40.1 | Improved |
| add CTC PLs to the CTC head in AR mode alongside attention PLs (ablation) | USR 2.0 · AVSpeech OOD AVSR | WER (%): 52.3 -> 40.1 | Improved |
| increase AR mode sampling probability toward 1 (ablation) | USR 2.0 · LRS3 ID vs AVSpeech OOD AVSR | WER (%) | Mixed |
| adaptive (linear or cosine) AR sampling schedule instead of constant 0.5 (ablation) | USR 2.0 · AVSpeech OOD AVSR | WER (%): 25.0 -> 24.8 (linear), 25.4 (cosine) | No clear effect |
| apply CTC merge and collapse to USR pseudo-labels alone (ablation) | USR · AVSpeech OOD AVSR | WER (%): 34.7 -> 36.9 | Hurt |
| CTC-driven teacher forcing instead of AR pseudo-labelling | USR · VSR training time | wall-clock training time: ~2x faster | Improved |
Closest papers by meaning
- Unified Speech Recognition: A Single Model for Auditory, Visual, and Audiovisual Inputs (NeurIPS 2024 · similarity 0.90)
- CR-CTC: Consistency regularization on CTC for improved speech recognition (ICLR 2025 · similarity 0.88)
- Uncertainty-Aware Self-Training for CTC-Based Automatic Speech Recognition (AAAI 2025 · similarity 0.86)
- Aligner-Encoders: Self-Attention Transformers Can Be Self-Transducers (NeurIPS 2024 · similarity 0.86)
- Align With Purpose: Optimize Desired Properties in CTC Models with a General Plug-and-Play Framework (ICLR 2024 · similarity 0.85)
- Self-Taught Recognizer: Toward Unsupervised Adaptation for Speech Foundation Models (NeurIPS 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.
