Designed Vocalizations Dataset powers AI voice conversion for monsters and robots
From animal calls to robotic voices—a new benchmark for non-human speech synthesis.
A team of researchers (Seolhee Lee, Minsu Kang, et al.) has introduced the Designed Vocalizations Dataset, accepted at InterSpeech 2026. While AI voice conversion has advanced rapidly for human speech, designed vocalizations—like monster growls, robotic tones, or alien sounds—remain underexplored due to a lack of public resources. This dataset addresses that gap by curating diverse raw vocal sources, including human speech and animal calls, then applying professional vocal effects processing to produce effect-modified variants. The result is a rich collection of non-human voices suitable for entertainment and media production.
The dataset includes a standardized test set with explicit seen/unseen splits over source timbre groups and preset styles, allowing researchers to assess generalization under controlled conditions. The team also provides baseline benchmark results to support reproducible evaluation. The dataset and demo samples are publicly available. By enabling systematic study of non-human voice conversion, this work paves the way for more realistic monsters in games, authentic robotic voices in films, and creative audio content in audiobooks and VR experiences.
- Curates raw human speech and animal vocalizations, then applies professional vocal effects to produce monster growls, robot voices, and more
- Includes a standardized test set with seen/unseen splits over source timbre groups and preset styles for generalization evaluation
- Accepted at InterSpeech 2026; dataset and demo samples are publicly available online
Why It Matters
Enables AI to generate realistic non-human voices for games, films, and audiobooks, filling a key gap in voice conversion research.