Visual features extracting & selecting for lipreading

This paper has put forward a way to select and extract visual features effectively for lipreading. These features come from both low-level and high-level, those are compensatory each other. There are 41 dimensional features to be used for recognition. Tested on a bimodal database AVCC which consists of sentences including all Chinese pronunciation, it achieves an accuracy of 87.8% from 84.1% for automatic speech recognition by lipreading assisting. It improves 19.5% accuracy from 31.7% to 51.2% for speakers dependent and improves 27.7% accuracy from 27.6% to 55.3% for speakers independent when speech recognition under noise conditions. And the paper has proves that visual speech information can reinforce the loss of acoustic information effectively by improving recognition rate from 10% to 30% various with the different amount of noises in speech signals in our system, the improving scope is higher than ASR system of IBM. And it performs better in noisy environments.