"tdmelodic", natural language processing software that automatically estimates Tokyo-style accents for Japanese speech synthesis, is now open source.

This software uses deep learning-based technology to estimate the accent of a word based on its surface form (the form in which the word appears in a sentence, such as kanji) and its reading (furigana).

This software can be applied to existing large-scale open-source Japanese dictionaries such as "NEologd" to automatically generate a large-scale accent dictionary for MeCab with a vocabulary size of several million words. The developer can start with the automatically generated dictionary using tdmelodic and NEologd as a baseline, and then improve the text-to-speech performance by correcting errors in the dictionary as necessary.

tdmelodic provides the following features
A function that takes the surface form and reading of a Japanese word as input and outputs the word's Tokyo-style accent.
Automatic generation of a large vocabulary accent dictionary for MeCab based on the existing Japanese morphological analysis dictionaries, UniDic and NEologd.

These results have been accepted to ICASSP2020, the top conference in the field of speech and acoustics.
Accent Estimation of Japanese Words from Their Surfaces and Romanizations for Building Large Vocabulary Accent Dictionaries.