Latest posts tagged with #visualspeechrecognition on Bluesky
Rethinking LLM Integration in Visual Speech Recognition
A Llama‑2‑13B model trained on combined LRS2+LRS3 reaches 24.7% word error rate (WER) on LRS3 and 47.0% on WildVSR; larger LLMs give only modest gains overall. Read more: getnews.me/rethinking-llm-integrati... #visualspeechrecognition #llm
GLip Framework Boosts Robust Visual Speech Recognition
GLip merges global facial context with local lip patches, boosting speech reading performance on LRS2, LRS3 and a Mandarin dataset despite lighting, occlusion and pose issues. Read more: getnews.me/glip-framework-boosts-ro... #glip #visualspeechrecognition
LLM Decoders Boost Lexical Accuracy in Visual Speech Recognition
Integrating Llama‑2‑13B with combined LRS2+LRS3 data yields a new VSR record: 24.7% WER on LRS3 and 47.0% on WildVSR. Scaling decoders alone gave modest gains. getnews.me/llm-decoders-boost-lexic... #llm #visualspeechrecognition #lipreading
