Breathy or Resonant - A Controlled and Curated Dataset for Phonation Mode Detection in Singing

This paper presents a new reference dataset of sustained, sung vowels with attached labels indicating the phonation mode. The dataset is intended for training computational models for automated phonation mode detection. Four phonation modes are distinguished by Johan Sundberg [15]: breathy, neutral, flow (or resonant) and pressed. The presented dataset consists of ca. 700 recordings of nine vowels from several languages, sung at various pitches in various phonation modes. The recorded sounds were produced by one female singer under controlled conditions, following recommendations by voice acoustics researchers. While datasets on phonation modes in speech exist, such resources for singing are not available. Our dataset closes this gap and offers researchers in various disciplines a reference and a training set. It will be made available online under Creative Commons license. Also, the format of the dataset is extensible. Further content additions and future support for the dataset are planned. 1. MOTIVATION: NARROW, WIDE, BREATHY, RESONANT SINGING IN VARIOUS