Non deterministic stochastic language models for speech recognition