Deep Reinforcement Learning for Articulatory Synthesis in a Vowel-to-Vowel Imitation Task

Area of Science:

• Speech Production Modeling
• Computational Linguistics
• Machine Learning

Background:

• Articulatory synthesis models human speech production.
• Existing methods often require extensive external training data.

Purpose of the Study:

• To develop a model-based algorithm for learning vocal tract control in articulatory synthesizers.
• To enable vowel-to-vowel imitation tasks without external datasets.
• To improve learning efficiency through simultaneous training of policy and dynamics models.

Main Methods:

• A model-based algorithm was developed to learn the control policy through interaction with a vocal tract model.
• Speech production dynamics were modeled and trained concurrently with the policy for enhanced sample efficiency.
• An acoustic word embedding (AWE) model was utilized for feature extraction and compact acoustic encoding.
• Early stopping was implemented to stabilize the training process.

Main Results:

• The proposed method successfully learned a vocal tract control policy without requiring external training data.
• Simultaneous training of the dynamics model and policy significantly improved learning efficiency.
• The integration of the AWE model was critical in guiding the policy towards a near-optimal solution.
• Extracted acoustic embeddings proved effective as inputs for both the policy and the dynamics model.

Conclusions:

• The developed algorithm offers a data-efficient approach to articulatory speech synthesis.
• Acoustic word embeddings are a valuable component for improving the performance of articulatory control policies.
• This method advances the field of speech synthesis by enabling more autonomous learning of vocal tract dynamics.