Persian Phoneme Recognition using Long Short-Term Memory Neural Network

Recently Recurrent Neural Networks (RNNs) have shown impressive performance in sequence classification tasks. In this paper we apply Long Short-Term Memory (LSTM) network on Persian phoneme recognition. For years Hidden Markov Model (HMM) was the dominant technique in speech recognition system but after introducing LSTM, RNNs outperformed HHM-based methods. We apply LSTM and deep LSTM on FARSDAT speech database and find that both LSTM and deep LSTM outperforms HMM in Persian phoneme recognition. Our evaluation show that deep LSTM achieves 17.55% error in FARSDAT phoneme recognition on test set which to our knowledge is the best recorded result.