Intoxicated Speech Detection using MFCC Feature Extraction and Vector Quantization

This study has been done on a technique which is suitable for tapping the telephonic conversation from a remote location to identify intoxication and consequent impaired brain activity that may cause criminal events e.g. DUI (driving under influence). This technique is time efficient, easy to use, non–invasive for the peoples and affordable for law enforcement personnel, bartenders/servers, court of law, coworkers/supervisors, clinicians, teachers and individuals who need to identify the presence and level of intoxication state in other peoples. The peaks in log Mel Filter Bank are main cues for identifying the sounds of speech. If a person is found drunk and his/her voice shows a great deal of variation, then this study describes an effective unsupervised method for query-by-audio sample speaker retrieval firstly by extracting MFCC features and then VQ (vector quantization) algorithms on the alcoholic audios. This method is also supported by verifying some speech parameters (fundamental frequency, jitter, shimmer). A set of twelve mel-frequency cepstrum coefficients computed every 10ms and which resulted the best performance i.e. 95% recognition with each of 8 speakers. The superior performance of the mel-frequency cepstrum coefficients may be an attributed to the fact that they better represent the perceptually relevant aspects of the short-terms speech spectrum.