A model will be presented that allows predicting binaural speech intelligibility in noise for normal-hearing (NH) and hearing-impaired (HI) listeners. The model inputs are the masker and target signals at the listener's ears, as well as their audiogram that is used to implement an internal noise (IN). The model computes per time frame and frequency band a binaural masking level difference (BMLD) when all signal levels are higher than the IN level, and the signal-to-noise ratio at the better ear (BE-SNR) using the maximum between the masker and IN levels. The BMLD and BE-SNR are integrated across time and frequency and summed to obtain a binaural ratio. The relative differences of the binaural ratios are compared to speech reception threshold (SRT) differences measured in listening tests. The model was validated on one experiment involving NH and HI listeners. Stimuli were anechoic and played using headphones. The target, in front of the listener, was presented simultaneously with two noise-vocoded speech maskers either collocated with the target or separated at +/- 90º. The noise was presented at different sensation levels and the relative target level was adapted to derive SRTs. The experimental data were accurately predicted by the model.