ZHG2P throws an exception when it encounters English input

[gbaian10]

from misaki import zh

g2p = zh.ZHG2P(version="1.1")
print(g2p("這是 example"))

Traceback (most recent call last):
  File "./example.py", line 4, in <module>
    print(g2p("這是 example"))
          ^^^^^^^^^^^^^^^^^^^
  File ".venv/lib/python3.11/site-packages/misaki/zh.py", line 80, in __call__
    segments.append(self.unk)
                    ^^^^^^^^
AttributeError: 'ZHG2P' object has no attribute 'unk'

For ZHG2P, if the text to be converted contains English, it will throw an exception.

How can we solve this problem elegantly?
A way to solve this issue is to preprocess the input by splitting it into segments.
Any other suggestions?

[hexgrad]

As of misaki >= 0.9.4 this should hopefully no longer crash, the unk symbol is ❓ by default.

To handle English, use en_callable as shown here: https://huggingface.co/hexgrad/Kokoro-82M-v1.1-zh/blob/main/samples/make_zh.py#L51-L70

[gbaian10]

Hi, thank you for the update. I have approximately two questions.

The first is about whether the __call__ method in ZHG2P, which currently uses a fixed pattern to split Chinese and English, can be adapted to allow users to pass in their own pattern for splitting these characters? (This argument would use the current pattern as the default value.)

The second question is regarding the fact that, before executing the regular expression for the Chinese-English split, the program has already converted all numbers into Chinese characters.
This causes issues where specific words cannot use their English pronunciation. For example, with "H2O," I would like it pronounced as "H two O" instead of "H 二 O."

---

Before receiving this update, I had already implemented my own Chinese-English segmentation, and the result is very similar to what would be produced by passing in the following en_callable.

en_g2p = en.G2P(trf=False, british=False, fallback=espeak.EspeakFallback(british=False))


def en_callable(text: str) -> str:
    return en_g2p(text)[0]

In Chinese, the symbol / seems to represent a word separator.
Does the presence or absence of this symbol have a significant impact on TTS, such as the sense of pause?
If this symbol is also added at the switch between Chinese and English, will it affect the synthesis result?
When I tried it, I didn't feel a big difference.