Researchers have developed DuET, a novel Dual-Encoder Transformer model designed for transcribing piano performances into musical scores. This model jointly learns note-level transcription and frame-level beat tracking by using bidirectional cross-attention between note and frame encoders. DuET also introduces TimeRoPE and Time-Pitch Rotary Positional Encoding to establish a shared performance-time reference and encode temporal and pitch relationships. Evaluations on the ASAP and Gangqinpu datasets show DuET improves score timing and notated duration, significantly reducing the MUSTER mean error on the Gangqinpu dataset. AI
IMPACT This research advances AI's capabilities in music information retrieval and symbolic music generation.
RANK_REASON The cluster contains a research paper detailing a new model and methodology for a specific task. [lever_c_demoted from research: ic=1 ai=1.0]
- arXiv
- Attention Sink Anchored Pruning
- Gangqinpu
- Hugging Face
- MUSTER
- Time-Pitch Rotary Positional Encoding
- TimeRoPE
- Transformer
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
