The discrete cosine transform (DCT) is widely used for image and video compression. Lossy algorithms such as JPEG, WebP, BPG and many others are based on it. Multiple modifications of DCT have been developed to improve its performance. One of them is adaptive DCT (ADCT) designed to deal with heterogeneous image structure and it may be found, for example, in the HEVC video codec. Adaptivity means that the image is divided into an uneven grid of squares: smaller ones retain information about details better, while larger squares are efficient for homogeneous backgrounds. The practical use of adaptive DCT algorithms is complicated by the lack of optimal threshold search algorithms for image partitioning procedures. In this paper, we propose a novel method for optimal threshold search in ADCT using a metric based on tonal distribution. We define two thresholds: pm, the threshold defining solid mean coloring, and ps, defining the quadtree fragment splitting. In our algorithm, the values of these thresholds are calculated via polynomial functions of the tonal distribution of a particular image or fragment. The polynomial coefficients are determined using the dedicated optimization procedure on the dataset containing images from the specific domain, urban road scenes in our case. In the experimental part of the study, we show that ADCT allows a higher compression ratio compared to non-adaptive DCT at the same level of quality loss, up to 66% for acceptable quality. The proposed algorithm may be used directly for image compression, or as a core of video compression framework in traffic-demanding applications, such as urban video surveillance systems.

This paper proposes a method for detecting the transformation of an audio signal using AI-based voice conversion (AI-VC) techniques applied to speech audio recordings (e.g., for the purpose of speaker impersonation or semantic content alteration), based on the analysis of acoustic and statistical features of the audio signal. The principles of operation of modern voice conversion systems are examined, with a particular focus on the Retrieval-based Voice Conversion (RVC) architecture. It is shown that the conversion process is multi-stage, involving the extraction of target voice features from the original speech recording – equivalent to vectorizing the speaker's behavioral characteristics – the construction of a target template (a voice synthesis model), the generation of the required semantic content, and the synthesis of a new speech audio recording that exhibits acoustically insignificant differences from the target. A review of informative features sensitive to such transformation and recognized as synthesis artifacts is conducted, including Mel-Frequency Cepstral Coefficients (MFCC), Linear-Frequency Cepstral Coefficients (LFCC), Constant Q Cepstral Coefficients (CQCC), statistical properties of the prediction error (Linear Predictive Coding, LPC), fundamental frequency contour and its variability (jitter, shimmer), spectral centroid and flux, and zero-crossing rate. A formal binary classification model and its software implementation based on Gaussian Mixture Models (GMM) are developed. Experimental validation of the proposed method on the ASVspoof2019 dataset demonstrated a classification accuracy of 93.9% (AUC = 0.939, EER = 13.52%). This indicates that the proposed method can provide protection for recipients of speech audio recordings against targeted vishing attacks of the spoofing type (voice identity fraud). The method can be implemented as a software module and integrated into messengers or corporate communication platforms. The relevance of this work is underscored by the absence of readily available products capable of detecting voice forgeries based on spectral analysis results, despite ongoing research efforts in both domestic and foreign institutions (achieving accuracy levels of 91–93%), as confirmed by the literature review.