This article is devoted to the development of voice authentication algorithm for access control in automated control systems. The existing methods of allocation of individual voice characteristics and construction of voice models are considered. The algorithm of voice authentication for voice control of the Internet of Things system in Russian on the basis of a neural network is offered. The peculiarity of the algorithm is the use of mel-frequency cepstral coefficients and the text independence of the voice message. Experiments aimed at identifying the optimal set of analysed parameters and evaluating the efficiency of the classifier and the authentication system as a whole are described.
In accordance with this, it is necessary to develop an access control system that includes authentication by voice characteristics without the use of code words or phrases. 2
Development of voice authentication algorithm The system of automatic user authentication by voice is being developed within the voice control system of Internet of things by PICT faculty of the ITMO University. The purpose of this work is to create the voice control system that can, on the basis of voice commands and data obtained from ontological descriptions of devices and indications of smart sensors, logically derive and generate scenarios for the smart (automated) system, for example, the smart home or the smart classroom. Voice control is based on local speech recognition.
This approach unlike cloud-based speech recognition has a number of advantages: 1. No issues related to availability, bandwidth and other factors that affect the speed of recognition inherent in cloud solutions. 2. Unlike cloud systems, there is an opportunity to configure the speech recognition system to solve a specific task. The quality of recognition depends on the language model used. In different application areas, different words have different probabilities. Cloud solutions use standard systems that use an average model of the language, or a model designed to solve the problems posed to the creators of the platform, which do not always coincide with the user tasks of the system. 3. Resource-efficient implementation of voice activation. To implement the activation function using a cloud system it is necessary to broadcast everything that the microphone records to the cloud in order to detect the passphrase. This leads to additional loading of the transmission channel and the cost of Internet traffic. 4. No additional financial costs. There are many open source (free) libraries and tools for local speech recognition, while cloud solutions are commercial and provide paid access to their services.
The following steps are implemented to solve the research task:
2. Preprocessing of the voice recordings. 3. Extraction of vectors of individual vocal signs. 4. Building the voice model based on voice characteristics. 5. Decision-making and verification.
The data is collected by the software for working with devices that record audio signals (microphones).
The VAD (Voice Activity Detection) algorithm based on the energy [
IF Ei < k
E; where k < 1;
Features are extracted from the pre-processed voice records. Feature extraction is implemented using the bob.ap library.
As a result 60-element vectors of mel-frequency kepstral coefficients (MFCC) [
In feature vector-based recognition, the Gaussian mixture model (GMM) [
Voxforge dataset [
Further, in the verification and decision-making step, the user is authenticated by comparing the received probability with the threshold value. The threshold value was determined by conducting experiments with speakers who did not take part of the speaker set G formation, known to the classifier (negative experiments), and speaker form the set G (positive experiments). From the obtained values of identification probability for negative and positive experiments, two Gaussian distributions - correct and erroneous identification were constructed for each user. The intersection point of these graphs of these distributions is the threshold value for a particular user of the system (Fig.3).
During the development of the voice authentication system, a number of experiments were conducted to identify the value of the VAD threshold coefficient. The choice of threshold factor is very important when working with energy-based VAD: too high value can cause cutting off frames that contain the speaker’s voice and, in turn, too low value can cause many non-vocalized or noisy fragments would not be excluded from the set of frames. Figure 4 shows the dependence of the classifier accuracy on the VAD threshold coefficient. Using VAD with the right threshold value can improve system performance by about 10% compared to raw data.
When testing the developed algorithm of voice authentication results with the following values (shown in figure 5): equal error rate EER = 7%, the coefficient of accurate verification 87.1%. 3
Conclusion This paper describes the analysis of methods of biometric authentication of the speaker by voice. The existing algorithms and methods of biometric authentication by voice, including text-dependent and text-independent algorithms are investigated. The analysis of the tools used in this field is also carried out, and the neural network with two hidden layers and the distribution of neurons 60:40:40:20 is selected. The result is a voice authentication algorithm for access control in automated systems. In addition, specialized software for voice biometric authentication system based on neural networks has been developed. The number of experiments on the selection of the VAD threshold coefficient was carried out. Such metrics as accuracy (87%) and EER (7.1%) were used to evaluate the system. Possible directions of development of this research are formulated.