Generative Adversarial Phonology: Modeling Unsupervised Phonetic and Phonological Learning With Neural Networks

Area of Science:

• Computational Linguistics
• Deep Learning
• Phonetics and Phonology

Background:

• Deep neural networks offer insights into learning internal representations from speech data.
• Generative Adversarial Networks (GANs) can model speech acquisition as a dependency between random space and generated data.
• Unsupervised learning in GANs, using raw acoustic data without language-specific assumptions, is suitable for phonetic and phonological learning.

Purpose of the Study:

• To propose a methodology for uncovering internal representations in GANs that correspond to phonetic and phonological properties.
• To investigate how GANs learn allophonic distributions in speech.
• To explore the manipulation of latent variables for controlling phonetic features in generated speech.

Main Methods:

• Training a Generative Adversarial Network on an English allophonic distribution (aspiration of voiceless stops).
• Developing a technique to identify latent variables corresponding to phonetic properties like the presence and spectral characteristics of 's'.
• Actively manipulating identified latent variables to control phonetic features in generated speech outputs.

Main Results:

• The trained GAN successfully learned the allophonic alternation, generating speech signals with the correct conditional distribution of aspiration duration.
• The methodology identified latent variables that approximate phonetic and phonological representations, such as the presence of 's' and its spectral properties.
• Manipulation of these latent variables demonstrated control over phonetic features (e.g., 's' presence, frication amplitude) in generated speech.

Conclusions:

• GANs can learn and represent phonetic and phonological information through latent variables, approximating linguistic representations.
• The learned dependencies extend beyond the training interval, enabling further exploration of representation learning.
• The findings provide a framework for interpreting how neural networks learn speech representations, with implications for understanding language acquisition and disorders.