CycleGAN-Based Representation-Level Translation of Asian-Accented English Speech Features

This exploratory study investigates whether Cycle-Consistent Generative Adversarial Networks (CycleGANs) can transform extracted speech representations from an Asian-accented English source domain toward an English target domain without paired recordings. The study uses 242 Asian-accented English recordings and 579 English recordings from the Speech Accent Archive and compares MFCC vectors, spectrogram images, and MFCC images. The original recordings were randomly partitioned into speaker-disjoint training, validation, and test sets before feature extraction and segmentation. CycleGAN training used the training partition, and auxiliary CNNs were used to quantify classifier-based target-domain assignment rather than direct accent-conversion accuracy. Supplementary CNN cross-validation used the combined training and validation representations, with the separate test partition excluded from fold construction. For Asian-to-English translation using MFCC images, 92% of the 100 evaluated generated representations were classified as English by the auxiliary CNN. This 92% value applies only to the Asian-to-English MFCC-image configuration and represents classifier-based target-domain reclassification. It does not directly measure perceptual accent reduction, intelligibility, waveform quality, linguistic-content preservation, or speaker-identity preservation. The MFCC-image configuration yielded the highest observed Asian-to-English target-domain reclassification rate under the evaluated settings; this does not establish intrinsic superiority of MFCC images. The results document configuration-specific target-domain assignments of CycleGAN-generated representations under the auxiliary classifier. Further work should include external evaluation, waveform reconstruction, ASR-based measures, listener studies, and balanced per-L1 analyses.