Robust estimation, interpretation and assessment of likelihood ratios in forensic speaker recognition