Discriminative Multi-Stream Postfilters Based on Deep Learning for Enhancing Statistical Parametric Speech Synthesis.

Area of Science:

• Speech synthesis
• Artificial intelligence
• Machine learning

Background:

• Statistical parametric speech synthesis using Hidden Markov Models (HMM) offers high intelligibility and efficiency for low-resource languages.
• However, HMM-based speech quality lags behind deep learning approaches.
• Postfiltering is a proposed method to improve HMM speech quality.

Purpose of the Study:

• To introduce a novel postfiltering technique for HMM-based synthesized voices.
• To enhance speech quality by applying discriminative postfilters based on long short-term memory (LSTM) deep neural networks.
• To specifically address distinct degradations in voiced and unvoiced sound segments.

Main Methods:

• Development of discriminative postfilters utilizing multiple LSTM deep neural networks.
• Modeling the mapping from synthesized to natural speech for voiced and unvoiced segments separately.
• Evaluation using five distinct voices, Mel cepstral distance, and subjective listening tests.

Main Results:

• Discriminative postfilters demonstrated significant advantages over standard HTS voices.
• The proposed method outperformed non-discriminative postfilters in speech quality enhancement.
• Objective and subjective evaluations confirmed the effectiveness of the approach.

Conclusions:

• Discriminative postfiltering with LSTMs offers a promising avenue for improving HMM-based speech synthesis.
• This technique effectively addresses quality degradation in synthesized speech, particularly for voiced/unvoiced segments.
• The approach provides a viable solution for enhancing artificial voices, especially in low-resource scenarios.