Comparative analysis of modality alignment algorithms in multimodal transformers for sound synthesis

• Published in: Sučasnij stan naukovih doslìdženʹ ta tehnologìj v promislovostì (Online) no. 2(32); pp. 49 - 57
• Main Authors: Mukhin, Vadym, Khablo, Yaroslav
• Format: Journal Article
• Language: English
• Published: 30.06.2025
• ISSN: 2522-9818; 2524-2296
• DOI: 10.30837/2522-9818.2025.2.049

Subject matter: this research focuses on the use of multimodal transformers for high-quality sound synthesis. By integrating heterogeneous data sources such as audio, text, images, and video, it aims to address the inherent challenges of accurate modality alignment. Goal: the primary goal is to conduct a comprehensive analysis of various modality alignment algorithms in order to assess their effectiveness, computational efficiency, and practical applicability in sound synthesis tasks. Tasks: the core tasks include investigating feature projection, contrastive learning, cross-attention mechanisms, and dynamic time warping for modality alignment; evaluating alignment accuracy, computational overhead, and robustness under diverse operational conditions; and benchmarking performance using standardized datasets and metrics such as Cross-Modal Retrieval Accuracy (CMRA), Mean Reciprocal Rank (MRR), and Normalized Discounted Cumulative Gain (NDCG). Methods: the study adopts both quantitative and qualitative approaches. Quantitative methods entail empirical evaluations of alignment precision and computational cost, whereas qualitative analysis focuses on the perceptual quality of synthesized audio. Standardized data preprocessing and evaluation protocols ensure reliability and reproducibility of the findings. Results: the analysis reveals that contrastive learning and cross-attention mechanisms achieve high alignment precision but demand considerable computational resources. Feature projection and dynamic time warping offer greater efficiency at the expense of some fine-grained detail. Hybrid approaches, combining the strengths of these methods, show potential for balanced performance across varied use cases. Conclusions: this research deepens understanding of how multimodal transformers can advance robust and efficient sound synthesis. By clarifying the benefits and limitations of each alignment strategy, it provides a foundation for developing adaptive systems that tailor alignment methods to specific data characteristics. Future work could extend these insights by exploring real-time applications and broadening the range of input modalities.