A nonlinear time alignment technique is presented in the framework of stochastic trajectory models (STM). We show how to obtain maximum likelihood (ML) estimates of model parameters, and how to use the technique during recognition with a slight additional computational overhead. Experimental results for a French 850 word continuous speech task are given. For a 10 speaker population, we test with various degrees of nonlinearity, and the introduced technique provides a slight improvement (about 1%) in the average word recognition rate.