Asmens, kalbančio skirtingomis kalbomis, atpažinimas pagal balsą

LTStraipsnyje apžvelgiami asmens atpažinimo pagal balsą metodai, taikomi teismo ekspertizėje, pateikiami automatinių metodų darbingumo vertinimo kriterijai bei pateikiamas asmens, kalbančio skirtingomis kalbomis, įtakos automatinio atpažinimo pagal balsą metodo tikslumui įvertinimas. Pateikiami kitos, ne gimtosios kalbos (rusų) bei įrašų trukmės įtakos identifikavimo tikslumui tyrimo rezultatai, gauti taikant Bajeso metriką. Taip pat apžvelgiama šių tyrimų eiga ir rezultatų vertinimas. [Iš leidinio]Reikšminiai žodžiai: Asmens identifikacija pagal balsą; Bajeso metrika; Balsų bazės; Kombinuotas metodas; Teismo ekspertizė; Bayesian metrics; Combined voice identification method; Forensic voice recognition; Voice database.

ENFor the world wide rapid spread of mobile application in the fixation of crimes today great attention is paid to person identification by voice in forensics. Crimes become international in the unified European area and criminal suspects are able to communicate with each other in several languages: Russian, Polish, English, etc. Auditory-linguistic features extraction, comparison and evaluation become problematic and only partial in phonoscopy forensic examinations because of the fact that identified person in investigative and comparative sound recordings speak in different languages. It should be noted that this part of analysis person identification by voice accuracy and quality is highly depends on expert (specialist) qualification, knowledge of not only his mother tongue, but also other languages. Currently in forensic practice (forensic voice comparison) is used semi-automatic personal identification by voice methods. In particular is performed auditory-linguistic analysis, when described general and specific personal voice and speech features, such as: voice height, voice clearness, timbre, speech rate, manner of speaking, intonation nature, articulatory phonetic sounds properties, etc. Acoustic-instrumental analysis is a special method of spectral pairs is used to find formants, their parameters and statistical calculate. Efficiency of automatic voice recognition systems in the best way can be evaluated using voice databases, collected using the records provided for phonoscopic examinations.Since the training voice base parameters have strong influence on the performance of the system and at the same time on the person identification by voice accuracy, in this article the efficiency evaluation criteria of automatic person identification by voice systems are reviewed and results of studies of such system efficiency using Bayesian metrics and various voice base parameters were presented. As shown by the results of tests performed automated voice searching system using the Log Likelihood Ratio (LLR) approach different languages have some influence on the recognition accuracy in the case if the same person speak different languages, such as comparative record - Lithuanian, investigative - Russian. Basically it depends on person skills and ability to speak other languages. Ratio of probabilities in all cases differs to the same person speaking Lithuanian and Russian. In all cases having probability ratio no sufficient, varies in the range from -5 and +3, additional analysis are necessary. It can be argued that LLR and identification accuracy highly depends on the length of the record. The record is longer, the person identification by voice is more correct and exact. Accuracy of application automatic person identification technology by voice depends on variety of factors: the system training level, the sound records length, the record quality, and other. So it should be noted that in forensic expertise or in phonoscopy (forensic voice comparison) the automatic person recognition method is used together with traditional speaker recognition by voice methods as one of the instrumental methods or a part of combined method. [From the publication]