A non-linear Bayesian filtering frame extended to extract single-channel ECGs as found in fetal ECG extraction from abdominal sensor. The recorded signals are modelled as the sum of several ECGs. Each of them is described by a dynamic nonlinear model, previously presented for the generation of a highly realistic synthetic ECG. Consequently, each ECG has a corresponding term in this model and can be discriminated efficiently, even if the waves overlap over time. The sensitivity analysis of the parameters for different values of noise level, amplitude and heart rate ratio between fetal and maternal ECG shows its efficacy for a large set of values of these parameters. This framework is also validated on fetal ECG extractions from actual abdominal recordings, as well as from actual twin magneto cardiograms.