P. Abdzadeh; H. Veisi
Abstract
Automatic Speaker Verification (ASV) systems have proven to bevulnerable to various types of presentation attacks, among whichLogical Access attacks are manufactured using voiceconversion and text-to-speech methods. In recent years, there has beenloads of work concentrating on synthetic speech detection, ...
Read More
Automatic Speaker Verification (ASV) systems have proven to bevulnerable to various types of presentation attacks, among whichLogical Access attacks are manufactured using voiceconversion and text-to-speech methods. In recent years, there has beenloads of work concentrating on synthetic speech detection, and with the arrival of deep learning-based methods and their success in various computer science fields, they have been a prevailing tool for this very task too. Most of the deep neural network-based techniques forsynthetic speech detection have employed the acoustic features basedon Short-Term Fourier Transform (STFT), which are extracted from theraw audio signal. However, lately, it has been discovered that the usageof Constant Q Transform's (CQT) spectrogram can be a beneficialasset both for performance improvement and processing power andtime reduction of a deep learning-based synthetic speech detection. In this work, we compare the usage of the CQT spectrogram and some most utilized STFT-based acoustic features. As lateral objectives, we consider improving the model's performance as much as we can using methods such as self-attention and one-class learning. Also, short-duration synthetic speech detection has been one of the lateral goals too. Finally, we see that the CQT spectrogram-based model not only outperforms the STFT-based acoustic feature extraction methods but also reduces the processing time and resources for detecting genuine speech from fake. Also, the CQT spectrogram-based model places wellamong the best works done on the LA subset of the ASVspoof 2019 dataset, especially in terms of Equal Error Rate.
M. Khanzadi; H. Veisi; R. Alinaghizade; Z. Soleymani
Abstract
One of the main problems in children with learning difficulties is the weakness of phonological awareness (PA) skills. In this regard, PA tests are used to evaluate this skill. Currently, this assessment is paper-based for the Persian language. To accelerate the process of the assessments and make it ...
Read More
One of the main problems in children with learning difficulties is the weakness of phonological awareness (PA) skills. In this regard, PA tests are used to evaluate this skill. Currently, this assessment is paper-based for the Persian language. To accelerate the process of the assessments and make it engaging for children, we propose a computer-based solution that is a comprehensive Persian phonological awareness assessment system implementing expressive and pointing tasks. For the expressive tasks, the solution is powered by recurrent neural network-based speech recognition systems. To this end, various recognition modules are implemented, including a phoneme recognition system for the phoneme segmentation task, a syllable recognition system for the syllable segmentation task, and a sub-word recognition system for three types of phoneme deletion tasks, including initial, middle, and final phoneme deletion. The recognition systems use bidirectional long short-term memory neural networks to construct acoustic models. To implement the recognition systems, we designed and collected Persian Kid’s Speech Corpus that is the largest in Persian for children’s speech. The accuracy rate for phoneme recognition was 85.5%, and for syllable recognition was 89.4%. The accuracy rates of the initial, middle, and final phoneme deletion were 96.76%, 98.21%, and 95.9%, respectively.
