[Bug] Tags < and > are removed during inference when text_frontend=True

[youngercloud]

Bug Report

Description

When setting text_frontend=True (or leaving it as the default), the < and > tags are removed from the text during inference.

For example:

INFO synthesis text 这也strong太strong离谱了吧！

---

To Reproduce

from cosyvoice.cli.cosyvoice import CosyVoice, CosyVoice2
from cosyvoice.utils.file_utils import load_wav
import torchaudio

# Initialize the CosyVoice2 model
cosyvoice = CosyVoice2(
    'pretrained_models/CosyVoice2-0.5B',
    load_jit=True,
    load_onnx=False,
    load_trt=False
)

audio_file_path = 'audio/48k.wav'
prompt_speech_16k = load_wav(audio_file_path, 16000)

for i, j in enumerate(cosyvoice.inference_cross_lingual(
    '这也<strong>太</strong>离谱了吧！',
    prompt_speech_16k,
    stream=False
)):
    torchaudio.save(
        'fine_grained_control_{}.wav'.format(i),
        j['tts_speech'],
        cosyvoice.sample_rate
    )

---

Expected Behavior

The inference should preserve the < and > tags as part of the input text.

这也<strong>太</strong>离谱了吧！

---

Actual Behavior

The tags < and > are removed, resulting in:

这也strong太strong离谱了吧

---

Environment

Please provide details about your environment:

• Operating System: WSL2
• Python Version: Python3.10 on conda

[darkacorn]

interesting .. maybe just a display thing as does work - on cross lingual that is

[youngercloud]

@darkacorn I reinstalled the Cosyvoice2 step by step on a new Ubuntu-based machine, and I got the log below.

2024-12-20 15:17:49,499 INFO synthesis text 这也<strong>太</strong>。

In terms of actual inference, the model uses the post-processed string, which indicates an incorrect string, and outputs the wrong audio.

The reason of re-installing is that I wanted to see if WeTextProcessing works properly, and it works fine for the above sentence.

Anyway, I will set the text_frontend to False and leave this issue open for a while to see if someone encounters a similar issue.