This paper proposes a corpus-driven spoken English evaluation system, which combines pronunciation feature extraction, SHO search and SVM rank determination to form a computational link. The database is built based on 2400 spoken language corpus of 300 learners, and the expert ratings are coded into four levels of A, B, C, and D. The system extracted MFCC, fundamental frequency, formant, energy, speech rate, pause ratio, stress offset and phoneme duration, and formed a 126-dimensional feature vector after standardization. Spotted hyenas optimizer selected subsets according to classification fitness, feature scale and redundancy, so that the feature dimension was reduced to 43. SVM classifier completed the grade label and score interval determination under optimized parameters. The 7:2:1 partition results show that SHO-SVM achieves 94.6% Accuracy, 93.8% Precision, 92.9% Recall and 93.3% F1-score, which are better than SVM, RF and BP network. The results show that the system has stable automatic pronunciation scoring ability and repeatable oral assessment ability, and can be applied to different scoring levels.
