Speech intelligibility estimation method from few samples in noisy conditions