Some Chinese text pronunciations are weird.

[lyris]

What happened?

[A bug happened!]
the .onnx version has unclear pronunciation of some words such as "质量", but the .pt version has no problem.
text: '图的质量总体来说比之前好，会有些图的质量非常高，但不稳定，并且画风会变来变去。'
phonemes: 'tʰu↗ tɤ ꭧɨ↘lja↘ŋ ʦʊ↓ŋtʰi↓ lai↗ʂwo→ pi↓ ꭧɨ→ʨʰjɛ↗n xau↓, xwei↘ jou↓ɕje→ tʰu↗ tɤ ꭧɨ↘lja↘ŋ fei→ꭧʰa↗ŋ kau→, ta↘n pu↘ wə↓nti↘ŋ. pi↘ŋʨʰje↓ xwa↘fə→ŋ xwei↘ pjɛ↘nlai↗pjɛ↘nʨʰy↘.'

example audios:
.onnx
.pt

Steps to reproduce

The latest onnx version kokoro-onnx 0.4.2

import soundfile as sf
from kokoro_onnx import Kokoro
from misaki.zh import ZHG2P

kokoro = Kokoro("kokoro-v1.0.onnx", "voices-v1.0.bin")
text = '图的质量总体来说比之前好，会有些图的质量非常高，但不稳定，并且画风会变来变去。'
lang = 'cmn'
g2p = ZHG2P()
for voice in kokoro.get_voices():
    if voice.startswith('z'):
        phonemes = g2p(text)
        print(f'{voice} speak {text} {phonemes}')
        samples, sample_rate = kokoro.create(phonemes, voice=voice, lang=lang, is_phonemes=True, trim=False)
        sf.write(f"output/onnx_version_{voice}.wav", samples, sample_rate)

native kokoro==0.7.12 misaki[zh]==0.7.12 (https://github.com/hexgrad/kokoro) has no problem:

import soundfile as sf
from kokoro import KPipeline

pipeline = KPipeline(lang_code='z', device='cpu')
text = '图的质量总体来说比之前好，会有些图的质量非常高，但不稳定，并且画风会变来变去。'
zh_voices = ['zf_xiaobei', 'zf_xiaoni', 'zf_xiaoxiao', 'zf_xiaoyi',
             'zm_yunjian', 'zm_yunxi', 'zm_yunxia', 'zm_yunyang']
for voice in zh_voices:
    for graphemes, phonemes, audio in pipeline(text, voice=voice):
        samples = audio.shape[0] if audio is not None else 0
        assert samples > 0, "No audio generated"
        print(f'{voice} speak {text} {phonemes}')
        sf.write(f'output/pt_version_{voice}.wav', audio, 24000)

example audio files are attatched above

### What OS are you seeing the problem on?

Window

### Package version

0.4.2

### Relevant log output

```shell

[thewh1teagle]

Try language.py example

[lyris]

The English has never been a problem, but there are issues with the Chinese in the example above. I modified it based on language.py, which uses an English example. The bug I reported is the Chinese example.

[fastfading]

same here on m1 mac

[fastfading]

@lyris
from deepseek

I'm not the expert , hope it can help you

if you fix it , could you send me the fix , thanks

[fastfading]

maybe some bug in
https://github.com/hexgrad/misaki/blob/main/misaki/zh.py