Forget DRML, you don’t need it: think Google Translate. There exists a corpus of audiobooks, with intonation (thanks to professional readers). There exists speech-to-text software. It should be possible to build a corpus of spoken phrases, correlated with the reader’s intonation, and then match text from a new work against the corpus to take a “best guess” approach to how to emphasize it. (Google is apparently doing something similar, using the UN’s gigantic corpus of multilingual translations of documents to build a database of phrase-level translations into multiple languages, which can then be applied to web pages. It’s effectively a non realtime mechanical turk, leveraging the historic work of an army of translators.
…
Posted by: Charlie Stross | February 27, 2009 2:39 PM
