A Generative Method for a Laryngeal Biosignal.

Area of Science:

• Bioacoustics
• Speech Science
• Computational Linguistics

Background:

• The Glottal Area Waveform (GAW) is crucial for quantitative clinical voice assessment, offering insights into vocal fold dynamics.
• Accurate modeling of GAWs is essential for advancing voice analysis and speech synthesis technologies.

Purpose of the Study:

• To introduce a novel method for generating synthetic Glottal Area Waveforms (GAWs) using Variational Autoencoders (VAEs).
• To demonstrate precise control over synthetic vocal fold dynamics via the Glottal Opening Vector (GlOVe).
• To explore the potential of synthetic GAWs in speech synthesis and phonetics research.

Main Methods:

• Utilized Variational Autoencoders (VAEs) to generate synthetic Glottal Area Waveforms (GAWs).
• Employed the Glottal Opening Vector (GlOVe) to manipulate the VAE latent space for controlling vocal fold closure and opening.
• Generated synthetic laryngeal biosignals with controllable oscillations, frequencies, and amplitudes.

Main Results:

• Achieved highly accurate synthetic laryngeal biosignals, with Normalized Mean Absolute Error values between 9.6 × 10⁻³ and 1.20 × 10⁻².
• Demonstrated significant training effectiveness, with reductions in key loss components up to approximately 89.52%.
• Successfully emulated realistic glottal opening changes and vocal fold oscillations.

Conclusions:

• The proposed VAE-based method with GlOVe provides a powerful tool for generating realistic synthetic GAWs.
• This approach offers precise control over laryngeal biosignal characteristics, aiding voice assessment and research.
• The findings suggest potential for developing advanced, natural-sounding speech technologies.