Gujarati Speech to English Text Conversion using attention based Sequence modeling

Real-time translation of spoken language to another language is crucial as it makes information accessible to people who do not speak the language. Translating spoken Gujarati into English text is a challenging task due to the linguistic and phonetic differences between the two languages. Traditional speech recognition systems often struggle with context and accuracy in low-resource languages such as Gujarati. This paper presents a comparative study of two deep learning approaches for real-time Gujarati speech-to-text conversion and translation: Simple Recurrent Neural Network (Simple RNN) with Attention and Bidirectional Long Short-Term Memory (BiLSTM) with Attention. The system processes Mel-frequency cepstral coefficients (MFCCs) extracted from audio inputs and applies both Simple RNN and BiLSTM layers with Attention to capture sequential dependencies while dynamically focusing on relevant speech frames. The models were trained using 80% from the data set, validated on 10% and tested on 10%, achieving training precision 100%, validation accuracy 100%, and test accuracy 100%. Furthermore, the integrated live translation module using Google Translate API achieved a fuzzy match score 84%, demonstrating the potential of the system in multilingual accessibility.