CNN and MKDE-Based Classification of Synthetic Speech Attribution

Speech synthesis algorithms have been continuously developed for years. They are easily available to the general public and have excellent performance. Therefore, if voice synthesis algorithms are used maliciously, they can be abused for various crimes such as fake news and fraud, including the impersonation of celebrities. Therefore, many studies have been conducted recently to solve problems caused by speech synthesis algorithms. As a result, various synthetic speech detectors have been developed in many previous studies. This paper proposes a synthetic speech classification model that can classify with which algorithms voice is synthesized in a situation where synthetic speech is affected by noise, reverberation, compression, etc. The proposed model consists of a Convolutional Neural Network (CNN) and a Multivariate Kernel Density Estimation (MKDE). The CNN model used the Mel spectrogram of the audio signal as a feature, and the MKDE model used the logit value of the CNN to estimate the Probability Density Function (PDF) of the training class. The data used for model learning and evaluation were provided in the 2022 IEEE Signal Processing Cup, where the final model performance shows an accuracy of 96.38% for the noiseless case and 95.5% for the noisy one, respectively.