Asymmetric and trial-dependent modeling: the contribution of LIA to SdSV Challenge Task 2