Answers are grounded in this paper and cite their sources.
Analyzes prosodic features in dialect speech, suggesting new methods for preserving linguistic diversity.
Japanese dialects are rapidly disappearing due to the decline of native speakers and ongoing linguistic changes, highlighting the need for efficient construction of dialect speech corpora to preserve linguistic resources. However, current corpus creation and analysis still rely heavily on expert knowledge, and unified analytical methods using speech processing technologies remain underexplored. This study focuses on building a model for analyzing regional differences in utterances of identical sentences. While our previous work demonstrated the effectiveness of joint training on automatic speech recognition (ASR) and dialect region classification using large-scale pre-trained models, it was limited to spontaneous speech, making it difficult to analyze prosodic features such as pitch accent differences across regions. This paper proposes a method for analyzing inter-regional prosodic features using read speech with identical content collected from dialect speakers across multiple regions. we construct dialect region classification models based on conventional acoustic features and self-supervised learning (SSL) models. We compute inter-regional speech similarity from learned representations and compare them with prosodic distances analyzed by linguists. Our results show that voice conversion–based data augmentation and multi-task learning improve dialect classification using SSL models. We also discuss the potential of these representations for analyzing regional prosodic variation.
No takes yet. Share an insight, caveat, or question.
Toyama et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: