Research map / Vision-language and multimodal / CLIP and zero-shot
Zero-shot keyword spotting for visual speech recognition in-the-wild
Themos Stafylakis, Georgios Tzimiropoulos
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 |
|---|---|---|---|
| add joint G2P decoder with auxiliary phoneme loss (ablation) | KWS network · LRS2 unseen keywords | EER: 8.46% -> 6.46% | Improved |
| use phoneme targets instead of grapheme targets in the auxiliary loss (ablation) | Joint KWS network · LRS2 unseen keywords | EER: 7.26% -> 6.46% | Improved |
| use phoneme targets instead of grapheme targets in the auxiliary loss (ablation) | Joint KWS network · LRS2 unseen keywords | FAR at MDR=1%: 40.26% -> 20.11% | Improved |
| train the encoder jointly with the KWS loss rather than only through the G2P decoder (ablation) | KWS network · LRS2 unseen keywords | EER: 7.22% -> 6.46% | Improved |
| train the encoder jointly with the KWS loss rather than only through the G2P decoder (ablation) | KWS network · LRS2 unseen keywords | FAR at MDR=1%: 30.99% -> 20.11% | Improved |
| train the encoder jointly with the KWS loss rather than only through the G2P decoder (ablation) | KWS network · LRS2 unseen keywords | Top-1 ranking rate: 34.14% -> 34.14% | No clear effect |
| use frame-level concatenation with BiLSTM stack instead of a whole-utterance video embedding | KWS network · LRS2 unseen keywords | EER: 32.09% -> 6.46% | Improved |
| add G2P decoder with auxiliary loss (ablation) | KWS network · LRS2 unseen keywords | Top-1 ranking rate: 23.71% -> 34.14% | Improved |
| add G2P decoder with auxiliary loss (ablation) | KWS network · LRS2 unseen keywords | Top-8 ranking rate: 55.90% -> 67.70% | Improved |
| use phoneme targets instead of grapheme targets in the auxiliary loss (ablation) | Joint KWS network · LRS2 unseen keywords | Top-1 ranking rate: 31.16% -> 34.14% | Improved |
| use the proposed visual KWS network instead of an ASR-based keyword search | KWS system · LRS2 unseen keywords | Top-8 ranking rate: 37.57% -> 67.70% | Improved |
| increase the minimum keyword length from 6 phonemes to 7 and 8 | Joint-G2P · LRS2 unseen keywords | error rates | Improved |
Closest papers by meaning
- Unified Speech Recognition: A Single Model for Auditory, Visual, and Audiovisual Inputs (NeurIPS 2024 · similarity 0.87)
- An Alternative Deep Feature Approach to Line Level Keyword Spotting (CVPR 2019 · similarity 0.87)
- Lip2Vec: Efficient and Robust Visual Speech Recognition via Latent-to-Latent Visual to Audio Representation Mapping (ICCV 2023 · similarity 0.87)
- VSRo-200: A Romanian Visual Speech Recognition Dataset for Studying Supervision and Multimodal Robustness (NeurIPS 2026 · similarity 0.84)
- Vision-Speech Models: Teaching Speech Models to Converse about Images (CVPR 2026 · similarity 0.84)
- MoME: Mixture of Matryoshka Experts for Audio-Visual Speech Recognition (NeurIPS 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.
