Advancements in Speech Recognition for Phonetically and Syntactically Similar Languages: Literature Review

This study explores advancements in Automatic Speech Recognition (ASR) and Natural Language Processing (NLP), emphasizing the role of deep learning models and specialized datasets. A comparative analysis of 15 research works highlights the use of various algorithms, from traditional statistical models to modern neural networks, in addressing linguistic challenges. Key findings include the enhancement of Arabic ASR through E2E transformer architectures, improving recognition, and reducing word ambiguity. RNN-based models achieved an 18.59% WER on a BC speech dataset, demonstrating improvements in speech recognition accuracy. 1D-CNN models effectively recognized emotions in Urdu, while ParsBERT outperformed other models in Persian NLP tasks. LSTM networks contributed to phoneme, syllable, and sub-word segmentation, and models like wav2vec2, DistilHubert, and Whisper were optimized for Galician ASR. The Arman-AV Persian dataset also enabled visual ASR with enhanced performance, and an LF-MMI-based TDNN-BLSTM system improved Urdu ASR accuracy. The creation of a phonetically balanced Arabic corpus enhanced Text-to-Speech synthesis. This study underscores the importance of dataset quality, model architecture, and computational optimization in advancing ASR, NLP, and SER. Future research should improve efficiency, reduce computational costs, and expand language resources for better multilingual performance.