Accent analysis based on formant frequencies of vowels in turkish language

Day by day, the crimes committed by physical means in the past give way to the crimes committed by electronic means with the increase of digitization. These crimes committed by electronic devices leave electronic evidences such as computers, mobil devices, network tools, storage devices behind. Examining audio evidences, mostly called as "audio forensic", is a sub-branch of digital forensic which contributes clarification of crimes by using electronic devices and data included in these devices. There are already plenty of audio analysis tools that can be used for audio forensic. These tools, which basically use DSP-based methods, are expected to automatically implement audio forensic tasks such as verification of audio integrity (editing and montage detection), spectrum analysis, speech enhancement (increasing the intelligibility of distorted records), speaker profile (age, sex, etc.) identification, analysis of audio masked by ambient noise, real-time audio analysis in judical tracking and identifcation of ethnic items like language and accent. In this work, speech production and phonetics, fundamental digital speech processing methods used for audio analysis, audio examinations in terms of digital forensics and technical and performance comparisions of audio analysis tools are covered in detail and a literature study has been done on accent identification which is one of the missing parts in these tools. After a literature search, it has been seen that no study has been performed for Turkish language on accent analysis until now. Accordingly, it is deduced that analysis based on vowel's formant frequencies is an important factor to classify accents in a language. Thus, in this thesis, we have performed a forensic accent analysis based on formant frequencies of vowels in Turkish language. In addition to detecting regional differences in formant frequencies via statistical analysis results, classification performance rates of up to 90% have also been obtained via classification analysis results. These findings, especially with trained classification algorithms using more number of speakers, will lead to the design of a simultaneous automatic accent analysis system based on the formant frequency values of vowels for the Turkish language in the future.