How to use Punctuation Models in FunASR

[chenmengzheAAA]

There are two types of punctuation models, offline and online.

1. The offline model, refer to punc offline for more details.

from modelscope.pipelines import pipeline
from modelscope.utils.constant import Tasks

inference_pipeline = pipeline(
    task=Tasks.punctuation,
    model='damo/punc_ct-transformer_zh-cn-common-vocab272727-pytorch'
)

rec_result = inference_pipeline(text_in='example/punc_example.txt')
print(rec_result)

The "text_in" supports three formats: raw inputs, text files, url files.

2. The online model, refer to punc online for more details.

Different from the offline model, the online model is used at each time of speech absence of voice activity detection(VAD) and
should be feed the cache extracted from the history text.

In the following code , we mark puctuation points on the pseudo streaming text where "|" represents the split points of VAD.

from modelscope.pipelines import pipeline
from modelscope.utils.constant import Tasks

inference_pipeline = pipeline(
    task=Tasks.punctuation,
    model='damo/punc_ct-transformer_zh-cn-common-vad_realtime-vocab272727'
)

inputs = "跨境河流是养育沿岸|人民的生命之源长期以来为帮助下游地区防灾减灾中方技术人员|在上游地区极为恶劣的自然条件下克服巨大困难甚至冒着生命危险|向印方提供汛期水文资料处理紧急事件中方重视印方在跨境河流问题上的关切|愿意进一步完善双方联合工作机制|凡是|中方能做的我们|都会去做而且会做得更好我请印度朋友们放心中国在上游的|任何开发利用都会经过科学|规划和论证兼顾上下游的利益"
vads = inputs.split("|")
rec_result_all="outputs:"
param_dict = {"cache": []}
for vad in vads:
    rec_result = inference_pipeline(text_in=vad, param_dict=param_dict)
    rec_result_all += rec_result['text']

print(rec_result_all)

[LauraGPT]

Update for current FunASR (main at 1d8080a, verified on 2026-07-15).

The original modelscope.pipeline examples above describe the ModelScope pipeline API. For current native FunASR code, the shortest path is AutoModel.

1. Offline punctuation restoration

Use ct-punc for the current Chinese/English model:

from funasr import AutoModel

model = AutoModel(
    model="ct-punc",
    device="cpu",  # use "cuda" when available
    disable_update=True,
)

text = "那今天的天气呢也是蛮好的啊你觉得怎么样呢我觉得还不错"
result = model.generate(input=text)[0]
print(result["text"])

Output from the environment listed below:

那今天的天气呢也是蛮好的啊，你觉得怎么样呢？我觉得还不错。

result also contains key and punc_array. The current aliases are:

• ct-punc -> iic/punc_ct-transformer_cn-en-common-vocab471067-large, for Chinese and English.
• ct-punc-c -> iic/punc_ct-transformer_zh-cn-common-vocab272727-pytorch, the older Chinese-only route.

Punctuation restoration is not text normalization. In particular, check product names and mixed-language tokens when exact spelling boundaries matter.

2. Online punctuation after VAD chunks

The online model needs the same dictionary object to be passed as cache for every VAD text chunk:

from funasr import AutoModel

model = AutoModel(
    model="iic/punc_ct-transformer_zh-cn-common-vad_realtime-vocab272727",
    device="cpu",
    disable_update=True,
)

chunks = [
    "那今天的天气呢也是蛮好的啊",
    "你觉得怎么样呢",
    "我觉得还不错",
]

cache = {}
pieces = []
for chunk in chunks:
    result = model.generate(input=chunk, cache=cache)[0]
    pieces.append(result["text"])
    print(repr(result["text"]))

print("".join(pieces))

Verified per-chunk output:

'那今天的天气呢也是蛮好的啊'
'，你觉得怎么样呢'
'？我觉得还不错'
那今天的天气呢也是蛮好的啊，你觉得怎么样呢？我觉得还不错

The comma or question mark can therefore appear at the start of the next returned piece. Append the pieces in order; do not strip their leading punctuation. Creating a fresh cache for each chunk produced no punctuation for the same three chunks.

The realtime model deliberately keeps the unstable tail open, so the final piece may not end with a period. At the final speech endpoint, run the completed transcript once through offline ct-punc if a fully punctuated final transcript is required. The online checkpoint is documented here.

3. Add punctuation to a Paraformer ASR pipeline

Paraformer does not emit punctuation natively, so attach the model at construction time:

from funasr import AutoModel

model = AutoModel(
    model="paraformer-zh",
    vad_model="fsmn-vad",
    punc_model="ct-punc",
)
result = model.generate(input="audio.wav", batch_size_s=300)

Do not add a separate punctuation model to SenseVoice, Fun-ASR-Nano, or Qwen3-ASR; those model families already emit punctuation natively. The current FunASR tutorial also has the concise AutoModel(model="ct-punc") example.

The raw-string paths above were tested with Python 3.12.3, FunASR 1.3.14, ModelScope 1.38.1, and PyTorch 2.11.0 CPU. Text-file and URL routing in the original post belongs to the ModelScope pipeline API and is separate from these native AutoModel examples.