Information processing device, information processing method, and program
By employing multitask learning to estimate intonation, accent phrases, and nuclei in Japanese text, the method improves accent estimation accuracy and speech synthesis quality by reflecting the hierarchical linguistic structure.
Patent Information
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- LINE WORKS CORP
- Filing Date
- 2022-07-28
- Publication Date
- 2026-05-25
AI Technical Summary
Existing accent estimation methods for Japanese text fail to adequately reflect the hierarchical language structure of accent phrases and nuclei, and do not account for the relationship with intonation phrases, leading to inaccuracies in accent estimation.
An information processing device and method that uses multitask learning to train a model for simultaneously estimating the delimiter positions of intonation phrases, accent phrases, and accent nuclei in Japanese text, employing a three-layer bidirectional long-short-term memory network model, conditional random field model, and autoregressive model to improve accuracy.
The proposed method enhances the accuracy of accent estimation in Japanese text by reflecting the hierarchical linguistic structure, resulting in improved speech synthesis quality and naturalness.
Smart Images

Figure 0007864585000014 
Figure 0007864585000015 
Figure 0007864585000016