Abstract. The paper introduces a new technique by augmenting MFCC features
with Machine Learning Models. The proposed technique is further integrated
with baseline TTS techniques to create a more robust deepfake speech detection
methods. Given the ways deepfakes have proven thusly convincing, it is much
more necessary to allow for trusted ways to distinguish them from natural speech.
This work uses popular TTS models like Tacotron2, Speedy Speech, and GlowTTS in the development of synthetic audio samples. The Mel-Frequency Cepstral
Coefficients for explored for the audio analysis. These features are then integrated
with Random Forest classifier to separate the natural speech from the fake audio.
Authors also designed a mathematical model that explains how the process is
carried out. The system was tested with a variety of audio files and produced great
accuracy in the detection of deepfakes with helpful insights on how this work
provides protection against fake audios.