A Study Of Forensic Speaker Comparison Based On Long-term Formant Distributions Between Mandarin And A Dialect

With the modernization of communication technology and audio-visual technology,the crime rate of telecommunication fraud,extortion,telephone threat and many other criminal or civil cases is increasing accordingly.LTF distributions,as a valuable measure to aid forensic speaker comparison,provide clear insight into the dimensions of a speaker's vocal tract and expose certain habitual characteristics that speakers have in speaking English,German and other languages.Its applicability for Forensic Speaker Comparison in certain dialect in China,such as Central Plain Mandarin dialect,or between Mandarin and the dialect,is still uncertain,and is under discussion in this study.Both quantitative and qualitative methods are employed in this study to analyze the features of LTF distributions.Firstly,Wavesurfer is used to cut the recordings in a way that only vowel sounds remained,and to output formant frequencies.Secondly,Matlab is used to display histograms based on formant frequencies from F1 to F4,and to do Fourier six rank curve fitting.Finally,the distinguishing power of LTF distributions for Forensic Speaker Comparison could be testified by comparing the features of LTF fitting curves,and comparing the Euclidean distance of LTF mean values calculated by SPSS,in the following three conditions: in Mandarin,in Central Plain Mandarin dialect,and between Mandarin and the dialect.The research finds: 1)fitting curves of LTF distributions(including peak values,number of peaks,kurtosis and skewness),and LTF mean values are helpful in inter-speaker discrimination.2)Generally,fitting curves of LTF distributions and LTF mean values tend to be stable within the same speaker but vary between different speakers based on detailed analysis.3)The LTF fitting curves and LTF mean values show robust discriminating power in Mandarin or the dialect,but the distinguishing power between Mandarin and the dialect is not as good as that in Mandarin or the dialect separately.Pearson product-moment-correlation shows that the correlation ofLTF mean values between Mandarin and the dialect is strong.4)Although LTF distributions are valuable for Forensic Speaker Comparison,due to the inherent flaws,failing to reveal the sound-by-sound variation,losing speech dynamics like transitions and co-articulations,LTF distributions need to integrate other acoustic and phonetic analysis methods to aid Forensic Speaker Comparison.Finally,Forensic Speaker Comparison based on LTF distributions is applicable in Mandarin,the dialect,and between Mandarin and the dialect.In spite of some limitations,this study broadens the application field of LTF distributions for Forensic Speaker Comparison.