The authors address the problem of developing efficient automatic speech recognition systems that maintain a balance between utilizing a wide knowledge of speech variability, while keeping the training manageable and improving speech recognition performance.