Realtime punctuation (CTTransformerStreaming) fails on empty text #3823

CTTransformerStreaming (funasr/models/ct_transformer_streaming/model.py), the class behind the realtime punctuation model iic/punc_ct-transformer_zh-cn-common-vad_realtime-vocab272727 used in examples/industrial_data_pretraining/ct_transformer_streaming/demo.py, fails on empty text. Since 4ad2a7b, the offline CTTransformer.inference returns text: "" and punc_array: None for empty or whitespace-only text, but the streaming class overrides inference without that check:

• With a fresh cache, "" or " " gives no words. tokenizer.encode([]) returns [], np.concatenate with the empty int32 cache ids turns the ids into float64, and the embedding raises the error users reported for the offline model in runtime error for audio file in which there is actually no speech (blank voice) #1386, 无人声音频时指定punc_model报错 #1625 and auto_model 中punc模型入参为空触发的bug #1660:

  RuntimeError: Expected tensor for argument #1 'indices' to have one of the following scalar types: Long, Int; but got torch.DoubleTensor instead (while checking arguments for embedding)
  

• When earlier calls left words in cache["pre_text"], an empty call runs the model on those words again but does not return them. If the model predicts no punctuation after the last of them, the output string is empty and if sentence_out[-1] in self.punc_list: raises IndexError: string index out of range.

AutoModel.generate(input="", cache=cache) passes the empty string straight to inference. In the demo's loop, a VAD segment with empty ASR text therefore raises the RuntimeError when it comes first, and can raise the IndexError later. runtime/python/websocket/funasr_wss_server.py only calls this model when len(text) > 0, but direct AutoModel callers have no such guard.

Repro on main (8710e1c) without downloading the model, using a tiny randomly initialised model of the same class:

from funasr.models.ct_transformer_streaming.model import CTTransformerStreaming
from funasr.tokenizer.char_tokenizer import CharTokenizer

model = CTTransformerStreaming(
    encoder="SANMVadEncoder",
    encoder_conf=dict(
        input_size=8, output_size=8, attention_heads=2, linear_units=16, num_blocks=2, input_layer="pe"
    ),
    vocab_size=5,
    punc_list=["<unk>", "_", "，", "。", "？", "、"],
    embed_unit=8,
    att_unit=8,
)
tokenizer = CharTokenizer(token_list=["<blank>", "<s>", "</s>", "<unk>", "是"])
print(model.inference([""], key=["utt"], tokenizer=tokenizer, cache={}, device="cpu"))

This raises the RuntimeError above. Expected: an empty result (text: "", punc_array: None), as CTTransformer returns, with the cache left as it was.

Environment: Windows 11, Python 3.11.17, torch 2.10.0+cpu, numpy 2.4.6, FunASR from source (pip install -e .).