A Robust Digit Recognition Model Research with Low Signal Noise Ratio

In this article, we absorbed the basic idea of DP in normalizing the temporal evolving characteristics of speech observation vector sequences, and designed a low SNR English digit recognition model based on whole MFCC sequences. Such model not only carries effectively the whole information but also normalizes the digit feature sequences by adjusting dynamically the frame length of each spoken digit, which can adapt the different speaking rates. By using frame time filter and multi sub-spectrum filter, noise interference to the digit can be partly reduced and the model´s adaptability to the backgrounds be improved. The experiment shows that the English digit error rate has reduced 30% at least, and 68.71% for best result after adding the new processing modules. In addition, the model is simple in structure and low in computation, and also easy to realize real time processing