Search

Recently people have difficulties distinguishing real speech from computer-generated speech so that the synthetic voice is getting closer to a natural-sounding voice, due to the advancements in deep learning and voice-generation techniques. This high level of naturalness allows for synthesizing a person target voice and reproduce someone's voice with high accuracy. Audio forensics is care with improving the speech intelligibility, enhancement, and analysis in order to prove the authentication of recorded audio files it is authentic and no manipulations were done to them. This work proposes the Audio verification model in Forensic Investigation using one Dimensional Convolutional Neural networks with new light architecture. The proposed model build a natural voice characteristics pattern by analysis the information available in an audio file and extract salient features in order to build a feature maps, and then tracked these feature maps over time is the same model to make the decision about whether the audio is Synthetic or not. The proposed model performed in two main stages: the pre-processing stage then verification predictor stage, each stage contained several steps performed different functions. The experiments were conducted on a last version benchmark dataset called the Fake-or-Real Dataset from the APTLY lab and ASVspoof 2019 dataset. The proposed model achieve successful results, so that the accuracy realizes 99.89% (± 1.78%), and the value of the loss function is 0.007 on the Fake-or-Real Dataset, and achieved a 0.006% Equal Error Rate (EER) on ASVSpoof 2019 dataset. more...

Nowadays in all everyday transactions, technological progress has become an intrinsic characteristic that depends on such electronic applications as financial and banking transfers, health care, project management, and other crucial life aspects.  The core of these applications is person Identification and/or verification steps which can be considered one of the complicated limitations.  Accordingly, the employment of biometric attributes can yield promising outcomes in these fields.  A One’s voice is a unique bio-feature whereby people can be authenticated and precludes others from assuming a one’s identity without their previous knowing or assent.  This work proposes a model with a new architecture to identify the person by exploiting the unique individual characteristics available in one’s voice based on deep learning. An augmentation method is utilized to increase the samples in the available dataset. The available temporal information at an input audio file is analysed then feature maps from this information are extracted which represent the salient temporal feature (time-domain features).  The decision is made based on tracking these voice features over time. Successful and promising results are achieved through this work, the accuracy is close to 99.81% (± 1.78%) and the values of loss function are close to 0.009 over VoxCeleb1 dataset for identifying 40 subjects. more...