Speaker recognition on compressed speech

We have investigated the influence of GSM speechcoding in the performance of a text-independent speakerrecognition system based on Gaussian Mixture Models(GMM) classifiers.The performance degradation due to the utilization ofthe three GSM speech coders was first assessed, usingthree transcoded databases, obtained by passing theTIMIT through each GSM coder / decoder. The codeddatabases were used for training and testing the speakeridentification system. The speaker recognition performance was also assessed using the original TIMIT and its8 kHz downsampled version.Then, different experiments aimed to explore featurecalculation directly from the encoded parameters, and tomeasure the degradation introduced by different aspectsof the coders were carried out.