Methods, systems, and computer readable media can be operable to facilitate controlled caching of text-to-speech data. When text is identified for a text-to-speech conversion, a duration value to be associated with the text may be determined, and the identified text and duration value may be included within a request for a conversion of the text. An intermediate server may retrieve a speech file that is generated in response to the conversion request, and the intermediate server may cache the speech file for a certain period of time that is indicated by the duration value.