In this paper, we compare the recognition performance which can be achieved for speaker independent isolated word recognition over the telephone line by standard phonemic Hidden Markov Models (HMMs) with a hybrid approach using HMMs together with a Multilayer Perceptron (MLP) to estimate the HMM emission probabilities. Recently, the latter approach has been shown particularly efficient for a large vocabulary, speaker independent, continuous speech recognition task (i.e. DARPA Resource Management database). Since this approach seems to be more robust for simple context-independent phoneme models, the aim of this paper is to compare the performance which can be achieved in the case of task independent training, i.e. when the phonemic models are trained on a database which does not contain the words used in the targeted application.