Search

Long-Term Formant distribution (LTF) is a relatively new method in forensic speaker comparison, by which the results have been proved to contain important speaker-specific information. However, few studies have been carried out for the fundamental issue that how long the speech sample should be collected. The current paper investigated the speech length threshold (SLT) by using Long-Term Cumulative Formants (LTCF) analysis, which was one of the LTF methods. The speech sample for each speaker was segmented into one-second length sub samples. Pearson's correlation coefficients were calculated for LTCF values of the whole speech sample and new set of speech samples that were formed by adding the immediately following sub sample onto the speech sample before it with a start from the first sub sample. The results show that SLT can be placed at about 70 seconds natural speech recordings (approximate 20 seconds only vocalic samples in duration), which are adequate to represent the whole vocal tract resonance characteristics.

Several ways have been proposed for forensic speaker identification based largely on formants, whose frequencies and dynamics can represent the individual vocal tract in a way. Since speakers now can use other languages in the process of crime, a new question that how to identify different forensic speaker reveals. Given the same recording materials of 3 different languages (Chinese, English and Korean), it is of great value to differentiate the specific speaker even if the language that speaker use is not his or her native language. In this paper, a new method for this case based on formant features and mathematical principles is presented. When talking about formant features, emergence and number of peaks, kurtosis and skewness, which are the most significant among F1, F2, F3 and F4 values, are first extracted for this case of our experiment. Through the comparison of correlations between those features based on long-term formant (LTF) distribution, the similarity and distinction of the three languages spoken by the same speaker can be visualized.