Software · AI singing
Warble: a choir in your pocket
Warble sings four-part hymns live, on the phone itself: any hymn, any verse, any key, any tempo, with no server and nothing to download per hymn. Built for Advent Hymnals, in collaboration with Brian Engineering.
16.6 MB
the whole singer, int8, downloaded once
4 voices
soprano, alto, tenor and bass, live, together or alone
2–3× real time
on a mid-range phone (Helio G99), with no dropouts
1–2 cents
pitch error: the notes come straight from the score
How it works
Nothing neural runs per sample
A 13 M-parameter network runs 86 times a second and predicts the colour of the voice. Fixed signal processing (a pulse train at the sung pitch and an inverse FFT) turns that into sound. That is why it fits a phone with room to spare.
The score drives the singing
Words become phonemes, notes become a pitch curve with natural scoops and vibrato, consonants land before the beat. Any hymn, any verse, any key or tempo, with no recording needed.
A cleaner voice by design
Compact vocoders that predict a phase for every frame leave an 86 Hz roughness. Warble predicts only magnitudes and builds a consistent source, so the roughness is gone, at about 0.5% extra compute.
Trained on real and converted singers
GTSinger's real singers, plus recorded hymns converted to real voices with permission and labelled exactly from the renderer that sang them. The result: +0.4 naturalness on hymns it never heard.
Where it sings
Advent Hymnals app
Live singers in the Android app's player, beside recorded singers: open any hymn and sing.
The web player
player.adventhymnals.org and every hymn page of adventhymnals.org, in the browser, with no plugin.
Choir practice
Hear your own part alone, at your tempo and in your key, with the other parts softer or silent.
Credits
Singing data: GTSinger (Zhang et al., NeurIPS 2024, CC BY-NC-SA 4.0); Warble's weights are published under the same licence and the singing stays free. Recorded voices: LUNAI Project (Katyusha, Keiro Revenant, Liam Thorne), used and converted with permission. Voice conversion: Seed-VC. Inference: onnxruntime.