The performance of ASR systems can degrade severely when there is a mismatch between die training and test conditions. This is the usual situation for ASR systems in real world appücations in which voice is corrupted by the presence of noise and it is not possible to obtain training data recorded under any possible acoustic condition. These circumstances gave way to extensive research on techniques aimed at providing ASR systems with a great robustness to these environmental differences.In this Ph. D. Thesis, the performance of a conventional ASR system facing three sources of distortion has been analysed: inter-speaker variability (as in any speakerindependent system), the distortion due to the transmission channel (as in most systems for remote Information access), and background noise (almost any ASR appücations must work under these conditions).Regarding inter-speaker variability, it has been researched the incorporation of múltiple acoustic modeling in a modular Large-Vocabulary ASR system for telephone-based appücations, making special emphasis in the improvement of the inclusión rate, without substantially increasing the computational and memory load.