Troubles with streaming decode #1724
Comments
|
would you mind posting the complete decoding command? Also, could you share part of the decoding results? We want to know the error patterns, e.g., more deletions or insertions? |
|
Hi, The complete decoding command for chunking mode is like as follows: When I use decode.py instead, the results are quite fine. But with this command the part of decoding results looks like as follows: %WER = 98.58 PER-UTT DETAILS: corr or (ref->hyp) Mostly, the recognition result is empty... Hope this helps |
|
Could you also show the same info about |
|
This is a result for modified beam search with beam=4, but results of a greedy decoding are quite similar: %WER = 25.40 PER-UTT DETAILS: corr or (ref->hyp) |
|
Would you mind sharing the complete decoding command? |
|
Please also post the first few lines of decoding logs, both for streaming_decode.py and decode.py The first few lines contain complete information about your environment and the command you are running. |
|
This is a log for streaming_decode run: |
|
This is a log for decode.py: |
|
Could you change the |
|
I tried different values but without any substantial effect. |
|
One of the values in except -1 |
|
This is a log of the run with chunk size 32: |
|
This is a recognition result: |
|
It seems I've found a reason: I use my own fbank features which are slightly different from torchaudio fbanks, but streaming_decode.py loads raw audio and computes fbanks on-the-fly. I've checked this hypothesis by just replacing fbank computation with a call to cut.load_features() and this worked almost the same as simulated streaming does! It would be great to add some flag into streaming_decode.py to check if pre-computed features are used in cuts |
|
It's great to hear that you fix it. By the way, I suggest you follow what we are doing in Icefall to extract features. It can save you a lot of time when you want to deploy your model with Sherpa. |
Hi,
I am trying to apply librispeech streaming recipe for my data (about 4K hours). I train zipformer with
--causal 1for about 30 epochs and tried to run both simulated_streaming (via decode.py) and chunking (via streaming_decode.py) with my testsets.While decode.py shows reasonable results (WER of about 30%), chunking version doesn't work properly showing multiple deletions and WER of over 98%).
It is worth noting that at earlier epochs WER is lower (about 80-85%) but later increases up to almost 100%.
Non-streaming version of the model (trained without --causal 1) also works well.
Please suggest what could be the source of such behaviour
The text was updated successfully, but these errors were encountered: