<p>English-speaking skill assessment is a crucial part of language learning, but conventional methods of assessment are usually time-consuming and subjective. The proposed study aims at introducing a new BERT English Condition Random Model to the GRU network (BERT(ECRM)-SSGRU) to use the model in scoring automated English speaking. The model is inclined towards the semantic and syntactic undertones of the English learning texts, driven by the situational embedding of the BERT-ECRM, and the information available at the GRU layer is sequentially interdependent to identify the linguistic flow and the coherence. A total of 5,000 samples of essay, article, and textbook passages have been annotated by hand with Positive, Neutral, and Negative sentiment as to the motivational or critical effect of the passage. This assessment was quite satisfactory since the data were separated into training (70%), validation (15%), and testing performance (15%) subsets. They used performance metrics such as Accuracy, Precision, Recall, and F1-Score to determine model efficacy. Comparative testing of the proposed BERT(ECRM)-SSGRU model against Logistic Regression, LSTM, and standard BERT indicates that the proposed model attains better performance with a total accuracy of 90.6% and balanced class-specific F1-Scores. The results can be visualized using semilogy, stacked bar, compass, and area plots, as well as a confusion matrix, to demonstrate that the model has been able to consistently improve its accuracy in the process of sentiment classification in English learning material. The results suggest that contextual embeddings applied together with sequential modelling provide a valid and unbiased system of automated assessment for English speakers. The potential of the educational practice is great compared to the proposed approach because the degree of subjectivity is lower, the degree of feedback is higher, and useful information is provided to affect the curriculum and support learners.</p>