Reproduce “Hierarchical Representation of Prosody for Statistical Speech Synthesis” 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
Antti Suni, Daniel Aalto, Martti Vainio
This research introduces an unsupervised, unified framework for hierarchically representing speech prosody, specifically prominences and boundaries, using continuous wavelet transform applied to fundamental frequency, energy, and duration. The method achieved performance on the Boston University Radio News Corpus that matched or exceeded existing supervised and unsupervised techniques for both detection tasks.
Key insights
- Prosodic prominences and boundaries can be effectively analyzed and represented within a single, unified computational framework, challenging traditional separate treatments.
- The combination of fundamental frequency, energy, and duration significantly improves the accuracy of both prominence and boundary detection, with duration playing a crucial role.
- Hierarchical time-frequency scale-space analysis using CWT provides an objective and robust method for capturing the multi-scale nature of prosody in an unsupervised manner.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.