Language Log

« previous post | next post »

Every year since 2005, speech synthesis researchers have organized a yearly Blizzard Challenge, "[i]n order to better understand and compare research techniques in building corpus-based speech synthesizers". Part of the research effort involves the general public, who are invited to perform a series of evaluations of the results.

Participation takes about one hour in total — but your participation is registered, so that you can leave at any point, and then return and take the evaluation up again at the point where you left off. If you're willing, please follow this link to enroll and participate.

The technology being evaluated is the kind of speech synthesis where new texts are synthesized by re-combining bits and pieces of recordings of real human voices, or perhaps by concatenating elements drawn from a statistical analysis of the same recordings. There are many different approaches to every aspect of this process: how to choose the bits and pieces, how to take them apart and put them together, how to modify the resulting combination (if at all), and so on. The central idea of the Blizzard Challenge series is to give everyone the same collection of recordings to start with, and the same texts to synthesize, and the same limited amount of time to get it all done, and then to compare the results so that everyone can learn more about how to improve the technology.

The name "Blizzard" arose because the original speech collections used came from the CMU ARCTIC databases. And the ARCTIC name was chosen partly because

… we chose to use out-of-copyright books from the Gutenberg Project. With most of these texts being at least 70 years old, we face the issue of language drift. The English language has changed considerably over the past centuries and we did not want to infuse in our prompt set archaic English sentences. Thus we have hand selected a set of short stories whose style is recognizably modern, if not completely contemporary. Partly for consistency and partly from personal preference, we selected stories largely from the early 20th century author Jack London. Many of these stories – famously “To Build a Fire” – depict the difficult living conditions of the Yukon. Other selected books also describe the far Canadian north, hence our moniker Arctic.

The 2005 challenge was based on. The voice building data for the 2012 challenge is

[a]udiobook data, segmented into utterances and with transcriptions […], comprising around 50 hours of speech material, of which around 32 hours have high-confidence transcriptions, with the remainder having transcriptions of lower confidence.

The audiobook recordings come from librivox.org, and consist of John Greenman reading four books by Mark Twain (A Tramp Abroad; Life on the Mississippi; The Adventures of Tom Sawyer; and The Man That Corrupted Hadleyburg, and Other Stories.

May 16, 2012 @ 8:33 pm · Filed by Mark Liberman under Computational linguistics

Permalink