| With the development of science and technology,the use of multimedia data in law enforcement and judicial fields is becoming more and more extensive.At the same time,because multimedia data such as audio and video are easily tampered with and forged,this has brought great trouble to law enforcement officers.In this case,as one of the important types of evidence,digital recording requires us to adopt some effective technologies to ensure its reliability in law enforcement and litigation.One of the main problems is analyzing the source of audio,i.e.determining which device a piece of audio came from.It has been found through research that due to the different physical properties of different microphone components,they leave unique and distinguishable traces in the recorded audio.Therefore,we can identify the microphone by analyzing the unique audio characteristics,and then match the microphone with the target device one by one,so as to achieve the purpose of audio source detection.The development of effective audio data analysis technology is of great significance for maintaining judicial justice and ensuring the authenticity of evidence.Audio source detection involves two subproblems of device identification and device verification.The device identification subproblem refers to analyzing a piece of target audio to determine which device among multiple candidate devices recorded it.The device verification sub-problem is to determine whether a target audio is recorded by a specified device.These two sub-problems are faced with both noise and non-noise situations,and the noise situations include speaker noise,environmental noise,and Gaussian white noise.Existing methods usually preprocess the target audio first,then extract some common audio features,such as Mel cepstral coefficients,power regularized cepstral coefficients,and linear predictive cepstral coefficients,etc.,and finally use machine learning methods to compare audio features.difference between them to obtain the recognition result.Although this type of method can achieve better results in non-noise conditions,the recognition accuracy will be greatly reduced in noisy environments.Therefore,it is an urgent problem to find new methods to improve the recognition accuracy of audio source detection methods under various noise conditions.The sub-problem of device identification is a multi-classification problem,while the sub-problem of device verification is a two-category problem.The traditional method also extracts audio features and then uses the classification method to identify them.It also faces a decline in recognition accuracy under noisy conditions.The problem.In summary,audio source detection is an important and challenging problem.In the face of different noise environments,it is necessary to develop new methods to improve the recognition accuracy and robustness of audio source detection technology.Audio source detection is an important technology that can detect and verify the authenticity,credibility,and security of audio data.Existing audio source detection methods have some problems,such as high complexity and low accuracy.In order to solve these problems,this paper carried out research work and proposed a series of improvement schemes,including two sub-problems of device identification and device verification.In the device identification subproblem,this paper proposes a residual network based approach for audio source detection.This method finds out the distribution characteristics of the audio samples by observing and analyzing the time-domain features and frequency-domain features of the audio samples.Through the analysis of this article,it is found that the traces left by components such as microphones will be widely distributed throughout the entire audio.Therefore,this paper proposes to use the short-time Fourier transform to convert the original audio from the time domain to the frequency domain to extract more microphone-related features,and finally classify the audio features through a convolutional neural network.At the same time,in order to improve the robustness of the model,this paper proposes a new multi-classification framework for device identification,which adds a noise classification output to the original microphone classification output.The experimental results show that increasing the noise classification output can effectively improve the recognition accuracy of the model,and the detection accuracy of the method proposed in this paper is higher than that of the existing audio source detection methods.In the device verification subproblem,this paper proposes an audio source detection method based on a structurally re-parameterized model.This method uses a multi-branch structure in the training phase to improve the accuracy of the model.In the reasoning test,it is converted into a single-branch structure through structural reparameterization,which not only has a higher reasoning speed,but also does not reduce the accuracy of the model.Experimental results show that the structural reparameterization model not only achieves good results in the equipment verification subproblem,but also achieves high accuracy in the equipment identification subproblem.To sum up,the two audio source detection methods proposed in this paper have achieved good results in the two sub-problems of device identification and device verification,and have high application value.These improvement schemes can not only improve the authenticity,credibility,and security of audio data,but also have a positive impact and significance on the fields of audio data analysis,identification,and verification. |
