Improving Phonetic Recognition with Sequence-length Standardized MFCC Features and Deep Bi-Directional LSTM

摘要

Phonetic recognition is one of the most challenging problems in the field of speech analysis. These applications can be mentioned such as dialect identification [1], mispronunciation detection [2], spoken document retrieval [3], and so on. There are different approaches to solve these problems such as improving the feature selection on input speech [4], applying deep learning technique [5] [6] [7] or combining both of them [8]. With the sequence data as the phonetics, the architecture which is based on recurrent neural network (RNN) is an appropriate approach [9]. It is even more powerful when combined with the improvement of features selection on input data. In our approach, we combine the Mel Frequency Cepstral Coefficients (MFCC) method with sequence-length to present the acoustic features of speech and use some RNN models to phonetic classification. Our experiments are implemented on the Texas Instruments Massachusetts Institute of Technology (TIMIT) [10] phone recognition dataset. Especially, our data processing and features selection method give consistently better results than other researches using the same neural network model. Currently, we have achieved the lowest error test rate (13.05%) by using Bidirectional LSTM, which is the best result in TIMIT dataset with the reduction of about 3.5% over the last best result [5] [6].