BERT-APC: Reference-Free Automatic Pitch Correction via Musical Context
The recently introduced BERT-APC framework offers automatic pitch correction without the need for reference pitches. It employs a stationary pitch predictor to determine the steady pitch of each note from detuned vocals. Subsequently, a context-aware note pitch predictor deduces the intended pitch sequence by utilizing a repurposed music language model. Additionally, a note-level correction algorithm rectifies inaccuracies while maintaining expressiveness. This method overcomes the shortcomings of current APC systems that depend on reference pitches or basic estimation techniques. The framework was detailed in a paper on arXiv (2511.20006) as part of a cross-replacement announcement.
Key facts
- BERT-APC is a reference-free automatic pitch correction framework.
- It uses a stationary pitch predictor to estimate pitch from detuned singing voice.
- A context-aware note pitch predictor infers intended pitch using a music language model.
- A note-level correction algorithm fixes pitch errors while maintaining naturalness.
- Existing APC systems rely on reference pitches or simple estimation algorithms.
- The paper is available on arXiv with ID 2511.20006.
- The announcement type is replace-cross.
- The framework aims to preserve expressiveness and naturalness in vocal performances.
Entities
Institutions
- arXiv