Singing Wikipedia (redux)

Singing Wikipedia was originally made as part of my PhD. I’ve appended this version with (redux) because it is an instance of a portfolio piece undergoing a significant rewrite to get it onto the web.

In the original version of Singing Wikipedia, to generate the choir voices I used a tried and trusted technique of phase-vocoding time-stretching (repurposing a Pd patch used in several other pieces). In this case I applied it to a sound file generated by text-to-speech (TTS). I was satisfied with that, but when I came to remake Singing Wikipedia it appeared that this route would not be possible within the Web Audio API. Amusingly, in a case of superseding myself, having already implemented Singing Wikipedia and moved on to resurrecting Protest Songs, I discovered that the simple voice engine eSpeak had actually been converted to JavaScript, so the old implementation was possible. Something of an artistic quandary, as I actually preferred this implementation of Singing Wikipedia, which I felt was closer to my original idea.

The genesis of the current implementation has something of a circuitous route. Attending a talk about Tom Whitwell’s Workshop System, the speaker was talking about the primitive onboard computer designed by Tom and the various scripts that users have contributed to the project. One of these was a ‘Speak and Spell’-style computer voice, implementing vocal synthesis derived from the original patents. Evidently I filed this nugget away in my brain because, when thinking about remaking Singing Wikipedia and referring to my original notes, I read myself saying that what I wanted was a computer ‘singing’. I wanted it to be reminiscent of Max Mathews’ early Bell Labs experiment in which the IBM mainframe ‘sang’ Daisy, Daisy (famously referenced in 2001: A Space Odyssey). I sort of reached this goal in a haphazard way by combining the primitive vocal synthesis of eSpeak with time-stretching, but it wasn’t really quite it. What I needed - and the Speak and Spell example told me - was some sort of formant synthesis to generate that primitive computer sings sound I was after.

I wasn’t starting from scratch: the original logic of the piece remained. A random article from Wikipedia is served up via their API. This article is sung by a choir whose fundamental pitch is derived from the number of Wikipedia articles per second since its birth in January 2001. This was a nod to John Luther Adams using the rotation of the earth as the fundamental in his piece The Place Where You Go to Listen (a seminal environmental sonification installation referenced many times in my PhD). The notes that the choir sing are derived from a pentatonic scale built on this fundamental and then chosen by the current article count (7,240,035 at the time of writing) as a gamut of intervals to sing - an example of my ‘weird numerology’. ‘Singers’ are given a range of bass, tenor, alto and soprano (by offsets), so they cover a range of roughly 2 octaves. Where the original simply rendered the whole article (the summary paragraph, to be exact) as TTS, this version extracts the vowel sounds as held notes (AA, OH, EE) and a selection of consonants that are ‘singable’ within the formant model, like R, L, M, N. I couldn’t get white-noise fricatives to sound good, ruling out S’s, X’s and so on, while K’s and other plosives also proved unsuitable. Sticking to these formant-generated phonemes gives the whole thing the ongoing drone, computer plain-chant, bucket-chemistry-Ligeti feel that I was after.

The visualisation was tricky - as I’ve written elsewhere, I don’t want the visual element to distract from a sound work, but I’m aware some visual feedback is what makes this a webpage. I experimented with a few versions of coloured circles and other shapes representing voices coming in and out. They looked great, but to my horror it was all a bit derivative of Brian Eno’s apps (it’s one thing having an influence, the other to rip them off wholesale!). Given the piece’s aesthetic, referencing primitive vocal synthesis on ancient mainframe computers, I built a mock-terminal display (which in retrospect is more Commodore 64 than IBM) and, inspired by the classic C64 ‘maze’ one-line code (10 PRINT CHR$(205.5+RND(1)); : GOTO 10), had the terminal produce rows of mysterious glyphs from Unicode characters. These glyph patterns relate back to the voice and the particular phoneme being sung and are meant to represent some machine language that we’re not privy to.

The original version of Singing Wikipedia is documented in my PhD thesis.