fasr-punc-ct-transformer 0.5.2

ct-transformer punctuation model for fasr

fasr-punc-ct-transformer

Chinese documentation

CT-Transformer punctuation restoration for fasr. Use it as the sentencizer stage after ASR to split raw recognized text into punctuated AudioSpan sentences.

Install

pip install fasr-punc-ct-transformer

Registered Model

Registry name	Class	Best for
ct_transformer	CTTransformerForPunc	Chinese and mixed Chinese-English punctuation restoration

The default checkpoint is iic/punc_ct-transformer_zh-cn-common-vocab272727-pytorch.

Pipeline Usage

from fasr import AudioPipeline

pipeline = (
    AudioPipeline()
    .add_pipe("detector", model="fsmn")
    .add_pipe("recognizer", model="paraformer")
    .add_pipe(
        "sentencizer",
        model="ct_transformer",
        disable_log=True,
        disable_pbar=True,
    )
)

Confection Config

[punc_model]
@punc_models = "ct_transformer"
disable_update = true
disable_log = true
disable_pbar = true

Inside a pipeline:

[pipeline]
@pipelines = "AudioPipeline.v1"
pipe_order = ["sentencizer"]

[pipeline.pipes]

[pipeline.pipes.sentencizer]
@pipes = "thread_pipe"

[pipeline.pipes.sentencizer.component]
@components = "sentencizer"

[pipeline.pipes.sentencizer.component.model]
@punc_models = "ct_transformer"
disable_update = true
disable_log = true
disable_pbar = true

Direct Model Usage

from fasr.config import registry

model = registry.punc_models.get("ct_transformer")()
sentences = model.restore("今天天气真好我想出去玩你觉得呢")
for sentence in sentences:
    print(sentence.text)

Use local weights:

model.load_checkpoint("/path/to/ct-transformer")

Parameters

Parameter	Type / range	Default	true	false	Change when
disable_update	bool	True	Skips FunASR checkpoint update checks	Lets FunASR check for updates	You need reproducible startup or want update checks
disable_log	bool	True	Suppresses backend logs	Shows backend logs	Debugging model loading or inference
disable_pbar	bool	True	Hides progress bars	Shows progress bars	Interactive scripts where progress output is useful

Generic checkpoint fields such as checkpoint, cache_dir, endpoint, revision, and force_download are inherited from the base model.

Notes

• restore(text) returns an AudioSpanList, not a plain string.
• Input text should already be recognized text. This plugin does not run ASR.
• For pipeline usage, put this model on the sentencizer component.

Dependencies

• fasr
• funasr
• Python 3.10-3.12