The increasing availability of music and video content through the internet and mobile platforms has led to a growing demand for technologies that support both content analysis and generation. In particular, users seek to extract symbolic musical info...
The increasing availability of music and video content through the internet and mobile platforms has led to a growing demand for technologies that support both content analysis and generation. In particular, users seek to extract symbolic musical information such as chords and melodies from audio for applications including score generation, music education, and personalized recommendations. Meanwhile, the rise of user-generated content has driven interest in automatically generating dance motions that align with given music.
This thesis presents a unified framework that integrates symbolic music recognition and music-driven dance generation using extracted musical features. By combining deep learning-based audio analysis with generative modeling, the proposed framework aims to automate labor-intensive tasks and lower the barrier to creative content production for non-experts.
First, an extended note-level singing melody transcription approach is proposed to directly predict the onset time, offset time, pitch, and note value of each note from raw audio. A pseudo-labeling strategy for note values, a token representation tailored for this task, and a set of comprehensive evaluation metrics are introduced. Experimental results demonstrate that the proposed method outperforms existing baselines and commercial software in terms of both quantitative evaluation and qualitative visualization.
Second, an event-based chord sequence modeling approach is introduced by reformulating chord recognition as a segment-level sequence-to-sequence prediction problem. This addresses the oversegmentation issues often seen in frame-level classification, where small fluctuations in prediction confidence can lead to frequent and unnatural chord changes. A Transformer encoder-decoder structure is employed, along with two types of token representations for time-aligned chord modeling and an encoder pre-training strategy based on chord similarity. The effectiveness of the proposed methods is validated by improved recognition accuracy and segmentation quality over conventional approaches.
Finally, this thesis proposes a music-driven dance generation model conditioned on the recognized melody and chord information. To effectively incorporate these symbolic musical features, the model architecture is modified to better encode the given music within music encoder. A non-autoregressive, diffusion-based model is employed to generate dance motions that are both physically plausible and musically coherent. The effectiveness of the proposed approach is validated through quantitative evaluations on motion quality, motion diversity, and music-motion alignment, as well as qualitative human evaluation. The results demonstrate that the generated motions better reflect both the temporal and semantic structure of the given music.