Roll Over Borges

Last week I came across an intriguing project by an MFA student at Yale, Gareth Long.  He started with an audio CD of the recent Edith Grossman translation of Don Quixote.  Then he processed the recording through speech recognition software trained to recognize the voice of the narrator, George Guidall.  The resulting text file was formatted and bound as a book. You can buy it here.

The “catch,” as it were, is that the product cannot be read as Don Quixote; imperfections inherent in the voice recognition software and in the training, and the fact that Guidall’s narration is not all rendered as “Guidall,” mean that the printed book is a distinct work of art — not a reproduction or derivative version either of Cervantes, or of Grossman’s translation – even though the binding gives it the appearance of being a copy of the translation.

At least so I conclude; judge for yourself.  The project website explains:

This book was generated by processing the Audiobook version of Don Quixote (Edith Grossman’s 2003 translation) through speech recognition software. For the software to work it must first be trained using a very particular script to understand an individual’s voice – in this instance, the narrator of the audiobook, George Guidall. To accomplish this, the words in the training script were found within the Grossman book version. These words were then located on the corresponding audio CD, removed from their contexts and rearranged as if George Guidall was speaking the training script. Certain words that were not found in the novel (‘module’, as an example, was not a common word in 17th century Spain) were created by combining words: the first syllable of ‘modesty’ was put with the word ‘jewel’ to create the phonetic sounds of ‘module’. After the audio files for the training speech were completed, this constructed speech reading was played to the speech recognition software, effectively training the computer to learn the voice of George Guidall. With this in place, the speech recognition software was ready to hear Guidall’s reading of Cervantes’ book.

The constructed voice of Guidall, however, did not make for a perfect cypher. Often, Guidall used different voices for characters or Guidall ran words together – not necessarily enunciating each word to its fullest, throwing off the accuracy of the software. At other times, words such as ‘Don Quixote’ were unknown to the software’s built-in dictionary and came out differently, in certain instances, as ‘Donkey vote.’ Further, much of the text reads like a Joycean stream, as the speech recognition software could not decipher the silent punctuation of the audio source. The moments in the text where punctuation does appear are most likely mis-hearings of other words (where ‘calm’ is heard as ‘comma,’ for example).

This generated text was then culled and printed as a book, closely resembling the standard paperback edition of the 2003 Grossman translation.

Beyond Don Quixote’s obvious canonical importance as the first modern novel, the story itself is about books and reading through plays with meta-narratives, questions of authorship, and the book as an object of fetish. Like the novel, this new version is also preoccupied with these and many other subjects, including the exploration of the differences in reception between the act of both reading and listening.

At the very least, this is a cool thing. 

There’s an image of the work below the jump.