Estimation of glottal flow waveform from speech requires recording of low-frequency sounds down to DC level. It causes a low-frequency noise or bias in the reconstructed glottal flow waveform. Removing this bias by linear highpass filtering (HPF) degrades the shape of glottal flow waveform. A nonlinear method based on empirical mode decomposition is proposed for removing the bias without using HPF while preserving the shape of the glottal flow waveform. The biased glottal flow waveform is decomposed into its intrinsic modes, then the low-frequency bias is estimated by using the higher modes. Glottal flow waveform is reconstructed by subtracting the bias in the time domain. The results show that the proposed method accurately estimates the bias and yields significantly better glottal flow waveforms than the conventional HPF method.