Research On Some Key Technologies Of Audio Evidence Authenticity In Noisy Environments

With the rapid development of modern speech processing technology, a large number of powerful audio editing software are constantly emerging. People can easily edit audio which cannot be easily resolved in the hearing. Audio evidence authenticity forensics has attracted more and more attention. Independent recording evidence must have three elements: authenticity, legitimacy and relevance. Authenticity is the foundation of legitimacy and relevance and also is the primary condition of recording evidence.This paper study the authenticity of the recording evidence and source of evidence from the perspective of signal processing in this paper. Their core content respectively contains tamper detection and recording devices identification. In order to make the decision better, we also research the improving forensics of recording evidence of which the core content includes speech enhancement. The main contribution of this dissertation is as follows:(1) As to there is no special characteristic parameters for recording device, This paper deeply studies the characteristic parameters of recording devices and raises three characteristic parameters of recording devices:low value in time domain, proportion and roughness. According to the frequency response characteristics of recording device, we modify the MFCC filter group, raise the cepstral coefficients of modified MFCC and combine various characteristic parameters as the mixed characteristic parameters with 116 dimension. Coexistence of devices features and speaker characteristics in the audio features has an influence on devices recognition and speaker recognition system. Facing this problem, we propose an oblique projection subspace tracking method of recording devices features and speaker characteristics, based on devices sub-feature space and speaker sub-feature space aren’t orthogonal.60 young men and women, each 10 different voice, recorded the voice with 11 different brand models of recording devices. This experimental analysis indicates that mixing characteristic parameters can effectively represent the characteristics of the recording devices and recognition rate rises about 10.4%than normal cepstrum parameters, recognition rate can be improved greatly, from 74.4%to more than 95.9%, by oblique projection subspace tracking method.(2) As to audio evidence contains electric network frequency characteristics, this paper focuses on the influence of non-stationary noise and interference that have never been considered in the past power grid frequency measurement and proposes a kind of high precision power grid frequency measurement method based on linear complete transformation. Theoretical analysis and experimental results show that this method can achieve 98.8% accuracy in all kinds of noise environment, having a certain anti-interference characteristics. So it can be well applied to the authenticity forensics of recording evidence research. The experimental results also show that this algorithm can effectively characterize the features of recording site. As to audio evidence does not contain electric network frequency characteristics, we separate the noise with the audio in the fractional cepstrum domain and check the authenticity of recording evidence by using the combined feature parameters. Focusing on analyzing the effect of jointing on the noise characteristics, this paper proposes a jointing frame unite feature detection algorithm based on fractional cepstrum transform. The experimental results show that fractional feature parameters of fractional cepstrum transform is much better than sliding window difference energy parameters proposed by Mailk. The tampering detection rate can reach 92% without noise and its correct recognition rate can reach 84.3% in non-stationary noise environment. So it has high practical value.(3) Forensic speech usually suffers from noise, distortion, interfering sounds, and other signal processing challenges that can impede proper analysis. This paper proposes a new speech enhancement method which adopts combined filter of magnitude and phase in fractional Fourier transform domain to remove amplitude spectrum and phase spectrum of background noise. Theoretical analysis and experiments are conducted and the experiment shows that it has significantly superior Signal-to-Noise ratio over the modified amplitude phase filtering method proposed by Hossain. It has wide application prospects in forensic speech recognition and speaker recognition.