AvashoG2P: A Multi-Module G2P Converter for Persian

The conversion of graphemes to phonemes (G2P) is a fundamental task in text-to-speech (TTS) and automatic speech recognition (ASR) systems. Over the years, G2P systems have evolved from rule-based and statistical methods to advanced neural network-based approaches. Despite these advancements, G2P conversion for Persian remains challenging due to the complex relationship between spelling and pronunciation and the scarcity of high-quality datasets. This paper introduces the AvashoG2P, a multi-module novel solution for Persian G2P conversion. The AvashoG2P system leverages a sequence-to-sequence (seq2seq) model with a GRU-based recurrent unit and an attention mechanism. This model is trained on both diacritized and non-diacritized words, enhancing its understanding of phonemes and their relationships. The system achieves a Word Error Rate (WER) of 15% and a Phoneme Error Rate (PER) of 5%, demonstrating its effectiveness. One of the critical components of AvashoG2P is its homograph disambiguation module, which utilizes a single model for all homographs, addressing a significant challenge in Persian text processing. Our method leverages a classification approach for homograph disambiguation, which assigns a phoneme label to the entire input window. Our system achieves high accuracy while optimizing for latency and memory consumption. We achieve significant improvements in accuracy and F1 scores using transformer-based models and machine learning classifiers. Our results highlight the superior performance of the XLMRoberta model among transformer models, with an F1 Weighted score of 94.7, and the SVC model among machine learning classifiers, with an F1 Weighted score of 89.96. Additionally, we present the AvashoG2P-Benchmark, a comprehensive test dataset designed to facilitate future research and benchmarking in Persian G2P tasks (available at: https://huggingface.co/datasets/PartAI/AvashoG2P-Benchmark).