Application of deep reinforcement learning to speech signal generation models to improve speech recognition performance

Deep reinforcement learning is a machine learning method that does not require supervised signals in training data. This method is particularly advantageous in scenarios where the preparation of a substantial volume of training data with supervised signals poses a significant challenge. However, it is challenging to construct a model with numerous output variations since it is necessary to explore the optimal parameters through iterative trial and error. Consequently, it has not yet been applied to speech signal generation models. In this paper, we propose a system that utilizes deep reinforcement learning to speech signal generation models. This system aims to convert input speech into speech that is more readily identifiable by speech recognition systems. Rewards are calculated for each frame based on phoneme recognition results. By introducing restrictions on the variation of conversion outputs, it prevents excessive conversion of the original speech and unstable learning. A performance evaluation was conducted using English speech spoken by a Japanese speaker, who demonstrated the highest word error rate (39.0%). The results demonstrated an improvement of approximately 1.1 points compared to the original speech.