how to use fst howtwords in python pipline?

[zyb8543d]

c++ decoding already supports hot words，how to use fst howtwords in python pipline?

[LauraGPT]

Sorry, we do not have this plan.

[LauraGPT]

Update for current FunASR: the 2023 reply predates the current Python hotword API.

There is one important distinction:

• The local Python AutoModel pipeline does not load a compiled WFST/FST graph.
• paraformer-zh supports SeACo neural hotword bias through the singular hotword argument.
• Weighted WFST/FST hotwords remain a C++/ONNX runtime feature.

For the current Python pipeline:

from funasr import AutoModel

model = AutoModel(
    model="paraformer-zh",
    vad_model="fsmn-vad",
    punc_model="ct-punc",
    device="cpu",
    disable_update=True,
)

wav_file = f"{model.model_path}/example/asr_example.wav"

plain = model.generate(
    input=wav_file,
    batch_size_s=300,
)[0]

biased = model.generate(
    input=wav_file,
    batch_size_s=300,
    hotword="魔搭",
)[0]

print(plain["text"])
print(biased["text"])

Verified with FunASR 1.3.14 on CPU:

欢迎大家来到么哒社区进行体验。
欢迎大家来到魔搭社区进行体验。

Both results contained the same 14 timestamps; only the biased term changed.

For multiple hotwords, use a whitespace-separated string:

model.generate(input=wav_file, hotword="达摩院 魔搭")

You can also pass a UTF-8 .txt file path, with one hotword per line:

达摩院
魔搭

model.generate(input=wav_file, hotword="/path/to/hotwords.txt")

The current parser for string, local-file, and URL inputs is here: generate_hotwords_list. The same paraformer-zh usage is in the Python tutorial.

For Fun-ASR-Nano, the API is different and uses the plural list argument:

model.generate(input=wav_file, hotwords=["达摩院", "魔搭"])

If you specifically require weighted WFST/FST decoding, use the C++/ONNX runtime or server and pass a weighted hotword file with --hotword:

阿里巴巴 20
魔搭 15

That path is documented in the online runtime guide and offline runtime guide. A Python websocket client may call that runtime server, but that is different from loading an FST into the local Python AutoModel pipeline.

For very large vocabularies, the tutorial also documents postprocess_hotwords; that is text-level correction after recognition, not decoder or WFST bias.