The effect of sampling variability on systems and individual speakers in likelihood ratio-based forensic voice comparison