MP3 audio compression can be undesirable in circumstances where high-quality music presentation is required and there is a lack of automated, evidenced, and open-source methods to determine this. This study introduced a new and accessible approach to discriminate between compression levels and identify lossy audio transcoding. Machine learning classifiers were trained on feature sets of audio analysis statistics, derived from multiple step-wise re-encodings of compressed audio samples. Two classifiers, a stacked model and a XGBoost-based model, had comparable accuracies to previous examples in the literature and marketplace (Stacked: 0.947, XGBoost: 0.970, Literature reference: 0.965, Commercial reference: 0.980). For transcoded samples, which hide compression levels with post-processing, the new classifiers were less accurate than existing methods. However, all methods were inaccurate in identifying transcodes where artificial noise was added via the µ-law encoder. A command-line implementation is available at gitlab.com/jammcfar/kbps_detect_proto.
(See document for exact affiliation information.)
https://aes2.org/publications/elibrary-page/?id=21671
(702KB)
Click to purchase paper as a non-member or login as an AES member. If your company or school subscribes to the E-Library then switch to the institutional version. If you are not an AES member Join the AES. If you need to check your member status, login to the Member Portal.
