Reproduce “Quality-Diversity Search in Sound Generation: Investigating Innovation Engines for Audio Exploration” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Björn Þór Jónsson, Çağrı Erdem, Stefano Fasciani, Kyrre Glette
An Innovation Engine architecture automates sound discovery by combining Quality-Diversity search with a pre-trained deep neural network classifier, allowing for the exploration and generation of diverse synthetic sounds. The system effectively navigates vast sonic spaces, creates specialized sounds across temporal durations, and overcomes the limitations of human evaluation in creative search.
Key insights
- Automated sound discovery through Quality-Diversity search guided by a discriminative model is a viable approach, yielding diverse and contextually relevant synthetic sounds.
- Co-evolution of CPPNs and DSP graphs is crucial for higher performance and subjectively more aesthetically appealing sounds, leading to simpler CPPN networks compared to CPPN-only approaches.
- The system leverages "evolutionary stepping stones," with ancestors of elite solutions frequently originating from different classes, including significant cross-contextual (e.g., musical to non-musical) transitions.
- Specialized multi-CPPNs for different frequency ranges can significantly reduce individual network complexity while maintaining or improving overall QD scores.
- Sounds generated demonstrate strong temporal specialization, indicating that a sound's 'elite' status for a given class is often duration-dependent, underscoring the importance of this dimension.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.