Our goal is to study the existing star formation rate calibrations based on emission-line luminosities and to provide new ones. We use the SDSS data release DR4, which gives star formation rates and emission-line luminosities of more than 100000 star-forming galaxies. We confirm that the best results are obtained with the Halpha calibration. This calibration has an uncertainty of 0.17 dex. We show that one has to check carefully the method used to derive the dust attenuation and to use the adequate calibration: in some cases, the standard scaling law has to be replaced by a more general power law. When data is corrected for dust attenuation but the Halpha emission line not observed, the use of the Hbeta emission line, has to be preferred to the [OII]3727 emission line. In the case of uncorrected data, the correction for dust attenuation can be assumed as a constant value but we show that such method leads to poor results, in terms of dispersion and residual slope. Self-consistent corrections, based e.g. on the absolute magnitude, give better results in terms of dispersion but still suffer from systematic shifts, and/or residual slopes. The best results with data not corrected for dust attenuation are obtained when using the observed [OII]3727 and Hbeta emission lines together. This calibration has an uncertainty of 0.23 dex.