Optimize English g2p function for improved speed and efficiency

[gordon0414]

Overview

This merge request optimizes the English g2p function to significantly improve its performance, resolving Issue #104. The primary changes include moving global variables to the top-level scope to avoid redundant initialization, simplifying conditions, and streamlining functions.

Changes:

• Global Variables:
  • Moved ARPA, _g2p, eng_dict, and tokenizer to the top-level scope to prevent redundant initialization.
• Function Optimization:
  • g2p Function:
    • Simplified conditions for words with apostrophes.
    • Refactored word-to-phone mapping logic for clarity.
  • __post_replace_ph Function:
    • Removed redundant checks and streamlined mapping replacement.
  • __refine_syllables Function:
    • Replaced explicit loop index with direct iteration.
  • __text_to_words Function:
    • Avoided re-initialization of the tokenizer by moving it to global scope.

Performance Impact:

These modifications resulted in a significant speed improvement for the g2p function
example test was conducted as inferencing a sentence 100 times sequentially

sentence = "In this paper, we propose 1 DSPGAN, a GAN-based universal vocoder."
g2p_results = [g2p(sentence) for _ in range 100)]

before modification

real 2m59.692s
user 3m8.842s
sys 2m1.637s

after modification

real	0m10.850s
user	0m12.223s
sys	0m24.916s

[tsukumijima]

@gordon0414
I was responsible for a major refactoring in Style-Bert-VITS v2.4.
I took a look at the pull request, but there were a few things I was concerned about.

With this change, the tokenizer loading process has been moved to the global scope. However, this is something to avoid.
bert_models.load_tokenizer() caches loaded tokenizers for each language and keeps them in memory until explicitly unloaded.
Also, (Style-)Bert-VITS2 has different tokenizers for each language.

Moving bert_models.load_tokenizer() to the global scope will load the English tokenizer the moment style_bert_vits2/nlp/english/g2p.py is imported.
Due to this behavior, in a use case where speech synthesis is performed only in Japanese, the originally unnecessary English tokenizer and his BERT pre-trained model are loaded and downloaded the moment the module is imported.

The current Style-Bert-VITS2 code imports style_bert_vits2/nlp/english/g2p.py only when performing English speech synthesis.
However, if you manually import the module separately, this may cause an impact.
In my opinion, heavy operations like loading BERT models/tokenizers should be able to be loaded selectively and at appropriate times as needed.

[gordon0414]

@tsukumijima I moved the loading tokenizer inside the __text_to_words function for the code integrity.

The improved performance is reserved!