In this paper, excellent results are provided in the calculation of likelihood-ratios in real forensic conditions within the bayesian framework for the evaluation of speech evidences with a GMM-based speaker recognition system. Reported experiments have been performed with speakers from the Ahumada/Gaudà database, where 249 (122 male and 127 female) acted as reference population for the evaluation of the intervariability in each speech evidence, and the remaining 30 multisession male speakers acted as true/false suspects. Different GMM models have been trained from telephone recording sessions with different selections of the test files simulating different real forensic conditions. Results are provided in the form of likelihood ratios (LR) and are summarized in the form of Tippet plots, which are used to validate LR-based systems. All reported experiments have been performed with IdentiVox software, a tool for forensic speaker recognition that is actually been tested with real cases at Guardia Civil labs.