LSTM-GTCC based Approach for Audio Spoof Detection
Sanil Joshi, Mohit Dua
Abstract
With the recent advancement in the internet and technology, the user’s private information is vulnerable to different threats. Similar to any biometric system such as: finger print, face recognition Automatic Speaker Verification (ASV) are also exposed to different spoofing attacks. In order to tackle different spoofing attacks, the ASV system has two main components. One is front end feature extraction part and other one is backend classification model. The main focus of the paper is to develop the less complex ASV system that is robust against voice spoofing attacks. The proposed work uses Gammatone Cepstral Coefficients (GTCC) feature extraction technique at front end and Long-Short Term Memory (LSTM) technique at back end. The proposed work is trained and developed in Voice Spoofing Detection Corpus (VSDC). This dataset contains single point replay attacks (OPR) and multi point replay attacks (1PR). The proposed work provides 2.38 %,2.72 % Equal Error Rate (EER) values in 1PR and 2PR attacks. Also, it provides 98%, 97.26% accuracy under same attacks. The work is also evaluated using validation accuracy and Area Under the Curve (AUC) values.