OpenJTalk Dictionary Bundling

HayaKoe's Japanese G2P (pronunciation conversion) depends on pyopenjtalk.

The original pyopenjtalk downloads its dictionary from the network on first import, but HayaKoe uses a fork that bundles the dictionary inside the wheel, eliminating this delay.

The Problem

The original pyopenjtalk, at import pyopenjtalk time, automatically downloads via HTTPS the OpenJTalk Japanese dictionary (open_jtalk_dic_utf_8-1.11, ~23 MB) if it is not present locally.

After download, it is cached at ~/.local/share/pyopenjtalk/ and is not re-downloaded afterward.

However, Docker containers start with an empty filesystem each time, so the download repeats every time a container is launched.

In environments where the network is blocked, the import itself fails.

Implementation

The custom fork (lemon-pyopenjtalk-prebuilt) was modified to include the dictionary files inside the package during wheel build.

• The build workflow downloads and extracts open_jtalk_dic_utf_8-1.11.tar.gz, placing it inside the pyopenjtalk/ package
• Dictionary inclusion is declared in pyproject.toml's package-data and MANIFEST.in

On the HayaKoe side, specifying lemon-pyopenjtalk-prebuilt as a dependency in pyproject.toml completes the integration.

After installation, you can verify the dictionary is included inside the package:

site-packages/pyopenjtalk/
  +- open_jtalk_dic_utf_8-1.11/   <- bundled dictionary
  +- openjtalk.cpython-310-x86_64-linux-gnu.so
  +- __init__.py

Improvement Results

• Network call on first import completely eliminated
• Works immediately after installation in offline and air-gapped environments
• Docker image build results are identical regardless of network conditions