<![CDATA[AI-generated music increasingly resembles human-produced songs, creating a need for reliable source detection in audio classification. This study aims to develop and evaluate a segment-level hybrid model that combines ConvNeXt with bidirectional long short-term memory (BiLSTM) to capture local spectral patterns and temporal relationships across a song. Each song is represented by 24 five-second segments. Every segment is converted into a 128 x 128 mel-spectrogram, processed by ConvNeXt to obtain 768 features, and arranged as a sequence for BiLSTM modeling. The base dataset contains 22,844 songs divided into 16,522 training, 877 validation, and 5,445 test songs. Under the same data split and training protocol, the proposed ConvNeXt-BiLSTM model achieved accuracy of 0.8408, balanced accuracy of 0.8450, sensitivity of 0.7002, and an F1-score of 0.8190. The ConvNeXt-only baseline achieved 0.8228, 0.8276, 0.6613, and 0.7934 for the same metrics. Fine-tuning used 3,200 additional balanced samples, while final evaluation used a separate balanced test set of 200 songs. With a calibrated threshold of 0.484, the fine-tuned model achieved an F1-score of 0.8556 and balanced accuracy of 0.8700, correctly identifying 97 non-AI songs and 77 AI-generated songs. These results show that temporal modeling across mel-spectrogram segments improves detection, although evaluation using additional music generators, codecs, recording qualities, and adaptive audio contexts remains necessary.]]>
Copyrights © 2026
