A method for data-driven lexical adaptation on the basis of a limited number of acoustic training tokens is discussed. The method is closely related to pronunciation modeling techniques. A set of pronunciation variants is generated by forced alignment, followed by a step to select promising pronunciation candidates by using a ranking function. The method has been validated on a database consisting of short utterances (proper names) spoken by native and non-native speakers. In the case of 5 training tokens per word, an improvement of 10-30 percent relative could be obtained compared to the baseline. A number of possible improvements of this method are discussed as well.