Time-Frequency Localization Using Deep Convolutional Maxout Neural Network in Persian Speech Recognition

In this paper, a CNN-based structure for the time-frequency localization of information is proposed Persian speech recognition. Research has shown that receptive fields’ spectrotemporal plasticity some neurons in mammals’ primary auditory cortex and midbrain makes facilities improve recognition performance. Over past few years, much work been done to localize ASR systems, using spatial or temporal immutability properties methods such as HMMs, TDNNs, CNNs, LSTM-RNNs. However, most these models have large parameter volumes are challenging train. For purpose, we presented called Time-Frequency Convolutional Maxout Neural Network (TFCMNN) which parallel time-domain frequency-domain 1D-CMNNs applied simultaneously independently spectrogram, then their outputs concatenated jointly fully connected network classification. To performance structure, used newly developed Dropout, maxout, weight normalization. Two sets experiments were designed implemented on FARSDAT dataset evaluate model compared conventional 1D-CMNN models. According experimental results, average score TFCMNN about 1.6% higher than addition, training time 17 h lower traditional Therefore, proven other sources, systems increases system accuracy speeds up process.