Speech Signal Splicing Detection system based on MFCC and DTW

One of the key forensics topics has been the detection of speech forgeries, mostly using real evidence in court. The transmission of digital speech recording data over several media exposes the data to the risk of being attacked or tampered with. Several people misuse the audio by altering that using editing software, such as Adobe, Audition CC, etc., which results in speech forgeries. So, to overcome these scenarios speech forgery detection method is deployed. A speech forgery detection method for splicing is implemented in this paper. Firstly, voiced segments are identified in the speech signal and calculated Mel Frequency Cepstral Coefficients (MFCC). These coefficients are considered as features and are stored in the database for the registered speakers. Similarly, features are calculated for each voiced segment of test signals and compared those features with the database by using dynamic time warping. This proposed method is tested on 225 original speech signals that are recorded in two different environments using two different microphones. By combining original recordings of two distinct speakers, a forged dataset of 4900 spliced speech signals is developed to test the efficacy of the developed method. An accuracy of 99.39% was attained and is superior to other existing methods.