Abstract The observation of the polarised emission from the Cosmic Microwave Background (CMB) from future ground-based and satellite-borne experiments holds the promise of indirectly detecting the elusive signal from primordial tensor fluctuations in the form of large-scale B -mode polarisation. Doing so, however, requires an accurate and robust separation of the signal from polarised Galactic foregrounds. We present a component separation method for multi-frequency CMB observations that combines some of the advantages of map-based and power-spectrum-based techniques, and which is direcly applicable to data in the presence of realistic foregrounds and instrumental noise. We demonstrate that the method is able to reduce the contamination from Galactic foregrounds below an equivalent tensor-to-scalar ratio r FG ≲ 5 × 10 -4 , as required for next-generation observatories, for a wide range of foreground models with varying degrees of complexity. This bias reduction is associated with a mild ∼20–30% increase in the final statistical uncertainties, and holds for large sky areas, and for experiments targeting both the reionisation and recombination bumps in the B -mode power spectrum.