AI Singers
Four simulated human throats read choral music from the sheet music and sing it in four parts: a 1547 anthem by Tallis and five chorales from Bach's Christmas Oratorio, in German. No recordings of singers are involved: every note is air pushed through a modelled larynx and vocal tract.
Thomas Tallis, If Ye Love Me
I sang in choirs for years, and I wanted to hear whether a computer could sing the way a chorister reads: from the page, in parts, breathing at the commas. The singers are built on CHIMERA, Jameson Hodge's open-source physical model of the human voice: a simulated airway with lungs, vocal folds, tongue and lips, run in real time in the browser. I added a reader that turns a public-domain score (MusicXML) into breath, pitch and lyrics for each voice, gave the four voices four different throats, and put each one on a 3D head so you can watch the mouths move as they sing.
The library opens with Tallis's four-part anthem, sung as written in the original key at the tempo printed in the score, and five chorales from Bach's Christmas Oratorio (BWV 248) sung in German: the score's own syllables go through a small German-to-phoneme front end, the lower voices follow the cantus's words, and the repeats are sung with their second verse. Every piece in the library names the edition its sheet music came from. Earlier I taught the same voices a barbershop quartet by fitting them to a real recording; more pieces will follow.
Listen and watch the singersPrivate preview behind an access code. If you'd like to hear it, introduce yourself and ask.
What I like about this
It is the same kind of work as my data projects: take a careful model someone else built, feed it a structured source, measure how far the output is from the truth, and close the gap. Here the model is a human airway, the source is a Renaissance score, and the truth is what a choir sounds like.
Credit where it is due: CHIMERA is the work of Jameson Hodge, released under the MIT license. The score is a public-domain transcription. The choral adaptation, the score reader and the singers' voices are mine.