Low-resource hidden Markov model speech recognition