Recent advancements Automatic Speech Recognition (ASR) methods that are based on ML algorithms have been remarkable, largely due to extensive hand-labelled dataset and advanced computing power, which have enabled the training of robust deep neural networks. The networks have achieved impressively low Word Error Rates (WER) on scholarly standards. However, systems often experience a decline in accuracy when dealing with noisy audio inputs. The research introduces enhancements to the noise resilience of the Efficient Conformer Connectionist Temporal Classification (CTC) framework by incorporating both auditory and visual data. Enhancements include refining lip- reading techniques with an Efficient Conformer back-end coupled with a ResNet-18 visual frontend, and the strategic placement of intermediate CTC losses. This approach leveraged Inter CTC remnant sections to moderate the rigid autonomy prerequisite of CTC systems. Furthermore, we have substituted the grouped attention mechanism of the Efficient Conformer with a more streamlined and effective method termed ‘patch attention’. The testing on the publicly accessible LRS2 and LRS3 lip-reading datasets demonstrates that the integration of audio-visual data not only improves speech recognition amidst environmental noise but also expedites the training process, achieving lower WERs in a quarter of the time. The Audio-Visual Efficient Conformer (AVEC) model we developed sets a new benchmark, attaining WERs of 0–2% on the LRS2 and LRS3 test. It has shown a great potential on Audio-Visual prediction rather than audio-only and visual only.