ICASSP · RESEARCH DEMO

Reference-Free Singing Pitch Correction via Music-Constrained Sequence Editing

Author(s) Name(s) · Author Affiliation(s)

A reference-free system that detects note-level pitch errors and edits them with singing-aware context and tonal constraints—without target melodies, accompaniment, or ground-truth keys.

Listen to examplesRead the paper

The listening section uses five F0-aligned, 8-second WAV comparison clips. Audio starts loading as soon as the sample cards appear.

Paper overview

Abstract

Original abstract

Index Terms—

Overview of the proposed reference-free singing pitch correction framework
Fig. 1. Overview of the proposed reference-free singing pitch correction framework.

Held-out real test set

Results against reference-free baselines

Input accuracy73.25%note-level exact pitch
Ours accuracy83.43%+10.18 points over input
Error repair46.81%inaccurate notes repaired
Harm rate3.19%correct notes changed
MethodAcc. ↑ΔAcc. ↑Repair ↑Harm ↓
Input73.250.00––
Key Snapping71.76-1.4920.769.62
BERT-APC79.44+6.1926.181.11
Ours83.43+10.1846.813.19
Acc. note-level exact pitch accuracyΔAcc. change relative to inputRepair erroneous notes repairedHarm correct notes changed

Reproduced from Table 1 of the paper. The proposed method improves accuracy by 3.99 percentage points over BERT-APC.

Qualitative examples

Listen to all reconstructions

Each group includes the original singing performance and three available reconstruction methods. WAV data is requested only when its Play button is pressed.