ONNX bug about icefall HOT 9 CLOSED

Completed.

from icefall.

Please tell us the exact command you are using and also please tell us the duration of your test wave.

from icefall.

The duration of test waves is 30sec.
I use following command for the decoding.

export PYTHONPATH='/workspace/inputs/alexg/asr/src/models/k2_2024/icefall/egs/tedlium3/ASR/zipformer:/workspace/inputs/alexg/asr/src/models/k2_2024/icefall:$PYTHONPATH'

export CUDA_VISIBLE_DEVICES="0"

python ./zipformer/onnx_pretrained.py
--encoder-model-filename zipformer/exp/encoder-epoch-50-avg-1.onnx
--decoder-model-filename zipformer/exp/decoder-epoch-50-avg-1.onnx
--joiner-model-filename zipformer/exp/joiner-epoch-50-avg-1.onnx
--tokens data/lang_bpe_500/tokens.txt
/workspace/inputs/alexg/asr/src/projects/en_eval/input/test_1.wav

from icefall.

Could you use a shorter wave, e.g., less than 10 seconds or 20 seconds ?

from icefall.

Great.
Well done.

For 10 seconds ONNX decoder works properly. Is it possible to add support to 30 sec or should I have to update my VAD so that it will not produce segments longer than 10sec?

from icefall.

Is it possible to add support to 30 sec

Yes, absolutely.

Please change

from icefall.

I changed the value from 1,000 to 10,000 and it did not help.

I found that maximal time that I can work with is 20sec. For 21sec I already have an error.

Maybe I should change this value during the training, maybe this value is saved somewhere in the model?

from icefall.

There must be some constant value about the length of some positional encoding vector in the code. You need to find and change it.

from icefall.

You are right. max_len currently represents time of 20 sec. It can be increased to 2,000 and then maximal time will be increase to 40 sec. But it is important to perform this change BEFORE exporting the ONNX model.

Thank you so much, this issue can be closed.

from icefall.