When a Cover Becomes a New Performance
YuE2's cover workflow begins with a composition, not with a waveform continuation. A source recording is first translated into a reviewed symbolic melody, then YuE2 uses that score together with new lyrics and a target style to create a new song. This chapter follows that audio-to-score bridge, explains why SheetSage2 and MERT2 play different roles, and uses the cover results to decide when full-score conditioning is worth the constraint and when melody-only conditioning is the better trade.
The crucial distinction is between preserving a musical condition and preserving a recording. YuE2 can receive a melody that came from an existing performance, but the public workflow does not promise the original singer, timbre, waveform, or sample-accurate continuation. It transfers selected musical structure into a new generative context. That is exactly where the score-centered design is useful—and where its limits must remain visible.
The cover starts by making audio legible
A recording is rich but difficult to edit through a text-and-style generator. It contains pitch, timing, voice identity, instrumentation, room sound, mixing, and performance nuance in one signal. YuE2's public generation interface does not take that waveform as a direct audio-reference condition. The documented bridge is to use SheetSage2 to transcribe the recording into an ABC melody, review the result, and then submit the score to YuE2 with lyrics and a new style.
SheetSage2 is therefore an upstream translator. It turns an audio performance into a symbolic object that can be inspected, corrected, and passed across the model boundary. The cover guide recommends a melody-oriented transcription for this purpose. The resulting score can retain vocal and instrumental melodic material while omitting chord symbols, giving YuE2 a stable melodic condition without forcing the source accompaniment into the new arrangement.
The transcription is not a neutral copy. It is an estimate made by another model, and errors can enter before YuE2 sees the request. A missed note can alter the contour the new singer is asked to follow. A wrong meter can shift the relationship between events and bar boundaries. A mistaken section order can make the new song structurally unlike the source even if a short phrase sounds familiar. Warnings, the score itself, meter, and section organization should therefore be reviewed before generation. A valid ABC file proves that the notation fits the supported form; it does not prove that the transcription captured the source performance correctly.
That review gate is where the layered model becomes practical. The source recording belongs to the evidence about what the original sounded like. The transcribed ABC belongs to the evidence about what YuE2 was asked to preserve. The generated audio belongs to the evidence about the new realization. If those layers are not saved separately, an error in transcription can be mistaken for a failure of YuE2, or a successful symbolic transfer can be mistaken for identity-preserving audio.
MERT2 sits nearby but does not replace the score. SheetSage2 loads the MERT-v2-FullSong encoder selected by its configuration. MERT2 itself produces continuous music representations for analysis; it is not the causal discrete semantic-token stream that YuE2 consumes during generation.