Persian sentences to phoneme sequences conversion based on recurrent neural networks

Abstract

Grapheme to phoneme conversion is one of the main subsystems of Text-to-Speech (TTS) systems. Converting sequence of written words to their corresponding phoneme sequences for the Persian language is more challenging than other languages; because in the standard orthography of this language the short vowels are omitted and the pronunciation ofwords depends on their positions in a sentence. Common approaches used in the Persian commercial TTS systems have several modules and complicated models for natural language processing and homograph disambiguation that make the implementation harder as well as reducing the overall precision of system. In this paper we define the grapheme-to-phoneme conversion as a sequential labeling problem; and use the modified Recurrent Neural Networks (RNN) to create a smart and integrated model for this purpose. The recurrent networks are modified to be bidirectional and equipped with Long-Short Term Memory (LSTM) blocks to acquire most of the past and future contextual information for decision making. The experiments conducted in this paper show that in addition to having a unified structure the bidirectional RNN-LSTM has a good performance in recognizing the pronunciation of the Persian sentences with the precision more than 98 percent.