S³ / AUDIO EXAMPLES / ICLR 2027

Zero-Shot Singing Style Transfer via a Structured Lyrics–Melody Bottleneck

We introduce S³, a score-conditioned, singer-prompted singing framework built around a structured lyrics–melody bottleneck.
S³ transcribes source audio with supplied lyrics into a symbolic score, then generates singing conditioned on a reference recording.

Singing voice conversion (SVC)

Choose a source song and a target voice, then play S³ and compare it with the source or other systems.

01

Choose an example

The song to convert

Target voice / reference

The recording that guides the generated voice

CURRENT EXAMPLELoading examples…Audio, baselines & score below
02

SVC output

The audio below is generated by S³ for the selected song and reference. You can switch to the source vocals or original mix at the same playback position.

space play / pause
S³ · Generated performance0:00 / 0:00

Baselines

Same selected song and reference

03

Transcription output

The piano roll and lyrics below show the predicted transcription. Click a note or syllable to seek to its position in the audio.

Piano roll
8s
Single-note syllable Melisma · one syllable, multiple notes

Note–lyric alignment

Click a syllable to seek; a timestamp to play.
Underlines mark melisma.

Timing follows the predicted score; sung timing may differ.Download MIDI ↗