Influence of GSM speech coding algorithms on the performance of text-independent speaker indentification

This paper investigates the influence, on theperformance of a text-independent speaker identification system, of the three speech coding algorithmsstandardized for use in the GSM wireless communication network. The speaker identification system isbased on Gaussian Mixture Models (GMM) classifiers. Only the influence of the speech codingalgorithms was taken into account. This was done bypassing the whole TIMIT database through eachcoding/decoding algorithm obtaining three transcodeddatabases. These databases were used for training andtesting the speaker identification system. Forcomparison purposes, the speaker identificationperformance was also assessed using the originalTIMIT and its 8 kHz downsampled version. <br><br> The results showed a significant performancedegradation when using the GSM transcodeddatabases, compared to the normal and downsampledversions of TIMIT. These results are in correspondence with the subjective speech quality of eachcoder.