UCSYSpoof: A Myanmar Language Dataset for Voice Spoofing Detection

Automatic Speaker Verification (ASV) is widely used in voice-based security mechanisms. It involves accepting or rejecting a person identification based on the individual's voice, a unique biometric feature. However, it faces many challenges and is vulnerable to direct or indirect attacks. Spoof voice detection is also an important component in secure voice authentication systems. Unfortunately, there is no spoof detection system using Myanmar language dataset. Spoof detection systems are important for many languages, including Myanmar, as they prevent fraud and misinformation, maintain trust, cultural and linguistic relevance, etc. Therefore, this paper proposes a Myanmar spoof voice dataset called UCSYSpoof, which contains both real and spoofed speech signals. End - to-end speech synthesis, vocoder-based speech reconstruction, and voice conversion techniques were used to generate the spoof speech based on 12,000 genuine speech signals. To demonstrate the impact of proposed dataset, a simple spoof detection model is implemented using long short-term memory (LSTM) and convolutional neural network (CNN) classifiers with linear frequency cepstral coefficients (LFCC) and Mel frequency cepstral coefficients (MFCC) features. Based on the empirical results, using CNN with LFCC and MFCC features achieves the comparable results on proposed dataset. The results show that the detection model has Fl-score of 0.99 and an equal error rate (EER) of 0.004, respectively.