SR-TTS: a rhyme-based end-to-end speech synthesis system

Area of Science:

• Artificial Intelligence
• Speech Synthesis Technology

Background:

• Deep learning has advanced text-to-speech (TTS) systems, but current models often lack naturalness and expressiveness.
• Slow inference speeds in conventional TTS systems reduce overall voice quality and efficiency.

Purpose of the Study:

• To introduce SynthRhythm-TTS (SR-TTS), an optimized Transformer-based structure for enhanced synthesized speech.
• To improve phonological quality, naturalness, and accelerate speech generation for increased inference efficiency.

Main Methods:

• SR-TTS utilizes an encoder, a rhythm coordinator with a pre-duration predictor, and a decoder.
• A self-attention-based feature predictor and causal convolution enhance speech naturalness, articulatory accuracy, and time-series consistency.

Main Results:

• Human evaluations demonstrate SR-TTS outperforms existing techniques in speech quality and naturalness.
• The model exhibits cross-linguistic capabilities, validated through training on English and Chinese corpora.

Conclusions:

• SR-TTS offers a significant improvement in synthesized speech quality and generation efficiency.
• The technology is well-suited for applications demanding high-quality, natural speech, including intelligent assistants and synthesized podcasts.