This paper introduces a research project that represents an innovative approach to e-learning applications targeting automatic feedback on the pronunciation of non-native speakers based on computer speech recognition (specifically for Czech). We have collected data from 187 speakers of different pronunciation levels from 36 languages, conducted a pilot project, and developed the first version of an attributive annotation system based on tagging isolated speech sounds. We briefly mention the results of this stage (especially the success rate of the trained model), which led us to change our strategy and move to the next phase of the development of the automatic speech recognition tool. In this article, we present the current and next project phases: the Anophone annotation tool, a new annotation system based on whole-word tagging (two- to four-syllable words). The result is a measurable improvement in both the model and the success rate of speech recognition.