funasr_onnx and funasr_torch SenseVoiceSmall ignore use_itn: use_itn=True still sends the woitn query #3859

SenseVoiceSmall.__call__ in runtime/python/onnxruntime/funasr_onnx/sensevoice_bin.py (line 155) and runtime/python/libtorch/funasr_torch/sensevoice_bin.py (line 150) reads only the textnorm keyword:

textnorm_input = kwargs.get("textnorm", "woitn")

use_itn is never read. So model(wav_or_scp, language="auto", use_itn=True) sends the woitn query (15), the same as use_itn=False. That is the call in runtime/python/onnxruntime/demo_sencevoice_small.py, runtime/python/libtorch/demo_sensevoice_small.py, examples/industrial_data_pretraining/sense_voice/demo_onnx.py and demo_libtorch.py, and in the ONNX and Libtorch sections of the SenseVoice README (also README_zh.md and README_ja.md). The README describes use_itn as "Whether the output result includes punctuation and inverse text normalization". SenseVoiceSmall.inference (AutoModel) sends the withitn query (14) for use_itn=True, as the C++ runtime does for svs_itn (default true).

Repro without model files, from the repository root with PYTHONPATH=runtime/python/onnxruntime (the same with funasr_torch and PYTHONPATH=runtime/python/libtorch):

from funasr_onnx import SenseVoiceSmall


class Stop(Exception):
    pass


def read_tags(language_input, textnorm_input):  # the first step of __call__
    print(textnorm_input)
    raise Stop


model = SenseVoiceSmall.__new__(SenseVoiceSmall)  # skip loading the model files
model.read_tags = read_tags
for kwargs in ({}, {"use_itn": False}, {"use_itn": True}, {"textnorm": "withitn"}):
    try:
        model("en.mp3", language="auto", **kwargs)
    except Stop:
        pass

Output on main (20f26b7), for both packages:

woitn
woitn
woitn
withitn

Expected: withitn on the third line (use_itn=True).

This is a regression from 925910c ("update onnx batch inference"). The use_itn=True call came with aa3fe1a and c5f7f11, which made both wrappers read use_itn like SenseVoiceSmall.inference; the next day, 925910c replaced that with the textnorm list/file reader and dropped use_itn. Passing textnorm="withitn" works around it.

Environment: Windows 11, Python 3.11.17, onnxruntime 1.23.2, torch 2.10.0+cpu; funasr_onnx and funasr_torch from runtime/python on main.