Improving Post-Filtering of Artificial Speech Using Pre-Trained LSTM Neural Networks

Area of Science:

• Speech Synthesis
• Deep Learning
• Signal Processing

Background:

• Deep learning post-filters aim to improve statistical parametric speech synthesis by mapping synthetic to natural speech.
• Long Short-term Memory (LSTM) networks are effective but have room for improvement in quality and efficiency.
• Current methods often use random initialization for LSTM networks in speech synthesis.

Purpose of the Study:

• To introduce a new pre-training approach for LSTM networks to enhance synthesized speech quality, particularly spectral characteristics.
• To improve the efficiency of speech synthesis enhancement processes.
• To evaluate the effectiveness of auto-associative pre-training for LSTM-based speech synthesis.

Main Methods:

• Implemented an auto-associative pre-training method for a single LSTM network.
• Utilized the pre-trained LSTM as an initialization strategy for post-filters in speech synthesis.
• Focused on enhancing Mel-Frequency Cepstral (MFC) parameters of synthetic speech.

Main Results:

• The proposed auto-associative pre-training initialization demonstrated advantages in enhancing MFC parameters.
• The pre-training approach achieved superior results in improving the statistical parametric speech spectrum compared to random initialization.
• Effectiveness was observed in most tested scenarios.

Conclusions:

• Auto-associative pre-training is a viable and effective method for initializing LSTMs in speech synthesis.
• This approach offers a more efficient way to enhance spectral quality in synthesized speech.
• The findings suggest a significant improvement over traditional random initialization techniques.