Visual speech recognition for small scale dataset using VGG16 convolution neural network

摘要

Visual speech recognition is a method that comprehends speech from speakers lip movements and the speech is validated only by the shape and lip movement. Implementation of this practice not only helps people with hearing impaired but also can be used for professional lip reading whose application can be seen in crime and forensics. It plays a crucial role in aforementioned domains, as normal person's speech will be converted to text. Here, it is proposed to enhance the visual speech recognition technique from the video. The dataset was created and the same was used for implementation and verification. The aim of the approach was to recognize words only from the lip movement using video in the absence of audio and this mostly helps to extract words from a video without audio that helps in forensic and crime analysis. The proposed method employs VGG16 pre trained Convolutional Neural Network architecture for classification and recognition of data. It was observed that the visual modality improves the performance of speech recognition system. Finally, the obtained results were compared with the Hahn Convolutional Neural Network architecture (HCNN). The accuracy of the recommended model is 76% in visual speech recognition.