Here we present an improvement of paleotemperature reconstructions for the Southern Ocean by combining new diatom data from the Pacific sector with published Atlantic and Indian sector reference data. The statistical analysis of 336 surface sediment samples recovered from a wide area of Southern Ocean environments defines a supra-regional reference data set for quantitative summer sea surface temperature (SSST) estimations. In situ temperature measurements covering the time span from approx. 1900 to 1991 were used as reference instead of more recent time series of satellite-derived data, possibly biased by ocean warming. Different transfer function (TF) models for the Imbrie and Kipp Method (IKM), the Modern Analog Technique (MAT), Weighted Averaging (WA), and Weighted Averaging Partial Least Squares (WAPLS) were tested. Best performance for IKM was obtained using the D336 set with 29 diatom taxa and three factors, resulting in root mean square error of prediction (RMSEP) of 0.833°C for SSST. MAT estimates were best with six analogs resulting in the lowest RMSEP of 0.812°C. WAPLS applied to D336 resulted in a RMSEP of 0.782°C. WA performed less well, expressed by a RMSEP of 0.974°C. Furthermore, two subsets for the Atlantic (D151) and the Pacific sectors (D107) were applied with IKM to test for the advantages of localized TFs. IKM-D151 and IKM-D107 performed comparably good as IKM-D336, with RMSEP of 0.71°C and 0.68°C, respectively. Application of the augmented reference data sets on two Pleistocene sediment records from the Atlantic (PS1768-8) and Pacific (PS58/271-1) sectors led to the best performance of IKM with D336, expressed by high overall communalities (>0.75) and fewer (PS1768-8) to no (PS58/271-1) no-analogs, compared to the regional data sets, proving IKM-D336 to deal better with higher assemblage variability. SSST estimates for both cores exhibit similar glacial/interglacial patterns for all four applied D336-based TF methods, with the best concordance between IKM and WAPLS.