Modulation spectrum-constrained trajectory error training for mixture density network-based speech synthesis.

Area of Science:

• Speech Synthesis
• Machine Learning
• Signal Processing

Background:

• Mixture density networks (MDNs) in statistical parametric speech synthesis address limitations of linear output layers, such as fixed variances and unimodal assumptions.
• However, MDNs lack the static-dynamic constraint crucial for high-quality speech synthesis training.

Purpose of the Study:

• To propose a novel training algorithm for MDNs that incorporates minimum trajectory error.
• To introduce a modulation spectrum-constrained loss function to mitigate the over-smoothing effect in speech synthesis.

Main Methods:

• Developed a training algorithm based on minimum trajectory error for MDNs.
• Implemented a modulation spectrum-constrained loss function during the training phase.

Main Results:

• The proposed training algorithm demonstrated significant improvements in objective performance measures.
• Subjective evaluations confirmed enhanced speech quality, indicating a meaningful improvement over existing methods.

Conclusions:

• The novel training approach effectively addresses the static-dynamic constraint limitation in MDNs for speech synthesis.
• The modulation spectrum-constrained loss function successfully alleviates over-smoothing, leading to higher quality synthesized speech.