Impact of Target Speech Length on Performance in Voice Conversion Modeling

In voice conversion (VC), which transforms acoustic characteristics of one voice (source speech) to another (target speech), it is essential to preserve the linguistic content of the source speech while accurately extracting style information from the target speech. However, in existing deep learning-based vc frameworks, introducing short target speech results in less reliable style extraction, leading to decreased performance. This paper presents a comprehensive analysis of the conversion performance of three state-of-the-art vc models with different network architectures, focusing on the impact of target speech duration. Our experiments demonstrate the impact of target speech length on speaker similarity and linguistic content intelli-gibility. Absolute scores and their variability differ depending on the specific neural frameworks of the vc models. These findings provide valuable insights for designing vc systems for applications with limited target speech duration.