Enhancing Spoken English Assessment Using Convolutional Recurrent Neural Networks

The assessment of spoken English is important to ensure accurate identification and evaluation of various factors that affect oral and written English. Conventional techniques or automated computational models, for instance, rule-based approaches, fail in challenges like the scalability in grading and the capacity to solve complicated tasks with a wide range of Speech accents and varieties. Such limitations call for better and quicker, accurate and suitable assessment approaches. In this present research, a new method in the evaluation of spoken English is proposed through the implementation of Convolutional Recurrent Neural Networks (CRNN). The CRNN integrates convolutional ones which learn spatial features of the sound signals and recurrent ones which address temporal characteristics of the speech, thus enabling the system to reasonably evaluate speech comprehensively. Applied in Python, the model works with features such as the Mel-spectrograms and responds to the pronunciation, fluency and grammatical aspects adequately. From the experimental evaluation, the use of the CRNN model improves on the previous ML techniques by obtaining high accuracy levels of 99.2%. This high performance shows the model for performing fast and efficient dynamic assessment of spoken English skills across the learners with different national languages and accents. The given method identifies and solves the problems of previous approaches, providing an accurate and fast way to assess the given language, with further research being directed at expanding the possibilities of the model's application to other multilingual datasets.