Introduce marine

[6gsn]

This PR introduces marine, the module for estimating accent using the DNN-based method.

Motivation

The priory motivation of this PR is to enhance pyopenjtalk by introducing marine, which is the Japanese accent module using the DNN-based method.

What I did

Currently, extract_fullcontext() supports the creating the full-context label from texts. So, I added the option to activate the estimation using marine and the processing for estimating the Japanese accents using marine. And then, the accents will be applied to the full-context label.

• before

  def extract_fullcontext(text):
      _, labels = run_frontend(text)
      return labels

• after

  def extract_fullcontext(text, run_marine=False):
      njd_features = run_frontend(text)
      if run_marine:
          njd_features = estimate_accent(njd_features)
      return make_label(njd_features)

To implement this modification, I also modified the existing feature as below;

• Separated the features in run_frontend() as run_frontend() and make_label() for linguistic feature extraction and for full-context label creation, respectively.
  • This modification helps make the processing simple whether the run_marine option is activated.

• Modified the data structure of the linguistic feature that be created by run_frontend() as below;

  • before

    In [1]: from pyopenjtalk import run_frontend
    
    In [2]: njd_result, _ = run_frontend("こんにちは")
    
    In [3]: njd_result
    Out[4]: ['こんにちは,感動詞,*,*,*,*,*,こんにちは,コンニチハ,コンニチワ,0/5,-1,-1']

  • after

    In [2]: njd_features = run_frontend("こんにちは")
    
    In [3]: njd_features
    Out[4]:
    [{'string': 'こんにちは',
    'pos': '感動詞',
    'pos_group1': '*',
    'pos_group2': '*',
    'pos_group3': '*',
    'ctype': '*',
    'cform': '*',
    'orig': 'こんにちは',
    'read': 'コンニチハ',
    'pron': 'コンニチワ',
    'acc': 0,
    'mora_size': 5,
    'chain_rule': '-1',
    'chain_flag': -1}]

  • This modification helps simplify the processing when the accent information by marine is applied with other linguistic features and creates the full-context labels.

Additional changes

• Implement njd2feature() function, which converts NJD Node to linguistic features defined as list[dict]
• Implement feature2njd() function, which converts linguistic features defined as list[dict] to NJD Node
• Add run_marine option to tts()
• Update README.md to describe the run_marine option

Todo

• [x] introduce marine
• [x] update interfaces
• [x] update test
• [x] update README

[6gsn]

Hi, @r9y9! I believe this PR is ready for getting review. May I ask you to approve the workflow and review this PR?

[6gsn]

@r9y9 Sorry for the dumb error 😭 I fixed the error and checked that testing for lint passed in my local environment (also testing for pytest as well).

[6gsn]

Hi, @r9y9 I updated the docs and README as well 👍