WIP: Add shallow fusion to C API

[csukuangfj]

trafficstars

Integrate changes from https://github.com/k2-fsa/sherpa-onnx/pull/147

TODOs

• [ ] Fix iOS demo
• [ ] Fix .Net demo
• [ ] Fix Python APIs

[kamirdin]

Hi, I'm using OnlineLMConfig by Python APIs, but it seams that it didn't work by set 'rnn lm onnx path' to 'model'，may I ask what should I do to using shallow fusion by Python APIs？

[csukuangfj]

but it seams that it didn't work by set 'rnn lm onnx path' to 'model'

How do you tell it does not work?

[kamirdin]

How do you tell it does not work?

By decoding a test set which has 1k samples, say 10k characters, but got exactly the same result compared to not using LM.

[kamirdin]

Before this, I obtained better results by using Python LM decode scripts from Icefall. This was with the same test set, ASR, and LM model. So, it was expected that better results would be achieved, even if only a few characters were changed.

[csukuangfj]

haracters, but got exactly the same result compared to not using LM.

How many lm scales have you tried?

[kamirdin]

python code ：

lm_config = OnlineLMConfig(
    model=lm,
    scale=scale,
)

print(lm_config)
print("="*30)

recognizer_config = OnlineRecognizerConfig(
    feat_config=feat_config,
    model_config=model_config,
    lm_config=lm_config,
    endpoint_config=endpoint_config,
    enable_endpoint=enable_endpoint_detection,
    decoding_method=decoding_method,
    max_active_paths=max_active_paths,
    context_score=context_score,
)
print(recognizer_config)

and than print out this:

OnlineLMConfig(model="base/with-state-epoch-21-avg-2.onnx", scale=1.1)
==============================
OnlineRecognizerConfig(feat_config=FeatureExtractorConfig(sampling_rate=16000, feature_dim=80), model_config=OnlineTransducerModelConfig(encoder_filename="./asr_model_chunk320/encoder.onnx", decoder_filename="./asr_model_chunk320/decoder.onnx", joiner_filename="./asr_model_chunk320/joiner.onnx", tokens="./asr_model_chunk320/tokens.txt", num_threads=8, provider="cpu", model_type="", debug=False), lm_config=OnlineLMConfig(model="", scale=0.5), endpoint_config=EndpointConfig(rule1=EndpointRule(must_contain_nonsilence=False, min_trailing_silence=2.4, min_utterance_length=0), rule2=EndpointRule(must_contain_nonsilence=True, min_trailing_silence=1.2, min_utterance_length=0), rule3=EndpointRule(must_contain_nonsilence=False, min_trailing_silence=0, min_utterance_length=20)), enable_endpoint=False, max_active_paths=4, context_score=1.5, decoding_method="modified_beam_search")

added .def_readwrite("lm_config", &PyClass::lm_config) at here: https://github.com/k2-fsa/sherpa-onnx/blob/36017d49c4f0b2f2f87feeeb0a40e54be4487b76/sherpa-onnx/python/csrc/online-recognizer.cc#L39 and rebuild, still print lm_config=OnlineLMConfig(model="", scale=0.5), of recognizer_config

[kamirdin]

by adding lm_config(lm_config), at https://github.com/k2-fsa/sherpa-onnx/blob/36017d49c4f0b2f2f87feeeb0a40e54be4487b76/sherpa-onnx/csrc/online-recognizer.h#L97C16-L97C16 It appears to be functioning, but it is encountering more missing errors compared to the Python script. It seems like there is still some work to be done on it. In any case, thank you for your assistance!

[csukuangfj]

https://github.com/k2-fsa/sherpa-onnx/blob/36017d49c4f0b2f2f87feeeb0a40e54be4487b76/sherpa-onnx/csrc/online-recognizer.h#L97C16-L97C16

Thank you for identifying the bug. Would you mind creating a PR to fix it?

[kamirdin]

Sure, I will create PR after more test finish

[rkjaran]

I'd like to use the online LM with the C API. What's the status on this?