There is a great concern regarding the misuse of deepfake speech technology to synthesize a real person’s voice. Therefore, developing speech-security systems capable of detecting deepfake speech remains paramount in safeguarding against such misuse. Although various speech features and methods have been proposed, their potential for distinguishing between genuine and deepfake speech remains unclear. Since speech-pathological features with deep learning are widely used to assess unnaturalness in disordered voices associated with voice-production mechanisms, we investigated the potential of eleven speech-pathological features for distinguishing between genuine and deepfake speech, i.e., jitter (three types), shimmer (four types), harmonics-to-noise ratio, cepstral-harmonics-to-noise ratio, normalized noise energy, and glottal-to-noise excitation ratio. This paper proposes a method of combining two models on the basis of two different dimensions of speech-pathological features to greatly improve the effectiveness of deepfake speech detection, along with mel-spectrogram features, to enhance detection efficiency. We evaluated the proposed method on the datasets of the Automatic Speaker Verification Spoofing and Countermeasures Challenges ASVspoof 2019 and 2021. The results indicate that the proposed method outperforms the baselines in terms of accuracy, recall, F1-score, and F2-score, achieving 95.06, 99.46, 97.30, and 98.59%, respectively, on the ASVspoof 2019 dataset. It also surpasses the baselines on the ASVspoof 2021 dataset in terms of recall, F1-score, F2-score, and equal error rate, achieving 99.96, 96.65, 98.18, and 15.97%, respectively.

The speech transmission index (STI) and room acoustic parameters (RAPs) are essential metrics for assessing speech quality and predicting listening difficulty in a sound field. Although STI and important RAPs, such as reverberation time and clarity, can be derived from the room impulse response (RIR), measuring the RIR in regularly occupied spaces is difficult. Hence, simultaneous blind estimation of STI and RAPs is an imperative challenge issue that must be addressed. However, most existing methods provide only a single parameter and require a massive dataset for model training. A deterministic method is presented for blindly estimating STI and five RAPs using a stochastic RIR model that approximates an unknown RIR. An algorithm is formulated that uses the temporal power envelope of a reverberant speech signal to determine the optimal parameters of the RIR model. A mathematical model of reverberation and dereverabation process was proposed is based on the temporal power envelope of the signals. This model maps the parameters of the RIR model to the observed reverberant signal. The estimated RIR can then be synthesized using the optimal parameters to estimate the STI and RAPs. A simulation was conducted to evaluate the simultaneous estimation of STI and five essential RAPs from observed reverberant speech signals, in comparison to the best existing previous work. The root-mean-square error (RMSE) and Pearson correlation coefficient between the estimated and measured values were used as evaluation metrics. In terms of STI, the proposed method achieves the best accuracy with an RMSE of 0.037. With regard to the reverberation time and other RAPs, the accuracy remains consistent with the previous works. The results show that the proposed method can effectively estimate STI and RAPs simultaneously without any training.

We have developed a robust scheme for blindly estimating the speech transmission index (STI) based on a convolutional neural network (CNN) with temporal amplitude envelope as features. When assessing the quality of acoustics in a room where there are people present, STI needs to be estimated without measuring the room impulse response (RIR) or using a modulation transfer function (MTF). This estimation can be problematic because a blind method based on the MTF has low accuracy when the stochastic models of RIR and the background noise are mismatched to real sound environments. We improve the accuracy of STI estimation in noisy reverberant spaces by using a CNN that takes the entire temporal amplitude envelope of an observed speech signal as its input. Simulations were performed to evaluate the proposed scheme and results showed that it can maintain the appropriate accuracy under various realistic room acoustic conditions with an average RMSE of 0.12 and correlation of 0.87. These results demonstrate that the proposed scheme can robustly and blindly estimate STIs in noisy reverberant environments.