Off-campus SJSU users: To download campus access theses, please use the following link to log into our proxy server with your SJSU library user name and PIN.

Publication Date

Spring 2026

Degree Type

Thesis - Campus Access Only

Degree Name

Master of Science (MS)

Department

Computer Engineering

Advisor

Magdalini Eirinaki; Bernardo Flores; Bertin Cordova Diba

Abstract

Automatic Music Transcription (AMT) for expressive piano performances remains a significant challenge due to the complexity of polyphony, fluctuating tempos (rubato), and shifting time signatures. While Recurrent Neural Networks (RNNs) have shown promise in predicting discrete musical tasks such as beat tracking, downbeat detection, and tempo estimation, they inherently struggle with long-range dependencies. Conversely, standard Transformer architectures, despite their success in global context modeling, often fail in rhythmic analysis due to their reliance on absolute positional encoding, which lacks a native understanding of relative musical distance. This thesis proposes a cascaded deep learning pipeline for piano transcription, replacing traditional sequential and vanilla self-attention bottlenecks with Rotary Positional Embeddings (RoPE). By implementing a RoFormer-based architecture, the model explicitly learns the relative timings between rhythmic events, resulting in more accurate downbeat and measure-level predictions. The proposed system processes acoustic features through a Convolutional Neural Network (CNN) before cascading through RoFormer blocks for beat, downbeat, and musical onset classification. Ablation studies conducted during this research demonstrate that standard Transformers struggle with beat tracking and quantisation performance compared to an RNN baseline. However, the introduction of RoPE natively helps to mitigate this limitation. The finalized RoFormer architecture achieves an F1-score of 89.6% on beat detection and 63.1% on downbeat detection, outperforming both the RNN and standard Transformer baselines. These results emphasize that relative positional awareness is a fundamental requirement for the successful application of self-attention mechanisms in musical time modeling.

Download

Share

COinS