Is there a way to check the audio that is actually used by the server side during decoding?
My configuration makes use of the sample librispeech yaml file provided (using gst-kaldi-nnet2-online), with models from our training. I compare the results of our evaluations using online2-wav-nnet2-latgen without rescoring and I'm getting a 2% discrepancy. But aside from this, some of the hypotheses from gst-kaldi-nnet2-online really tell me that the server seems to be corrupting the audio files.
Is there a way to check the audio that is actually used by the server side during decoding?
My configuration makes use of the sample librispeech yaml file provided (using gst-kaldi-nnet2-online), with models from our training. I compare the results of our evaluations using online2-wav-nnet2-latgen without rescoring and I'm getting a 2% discrepancy. But aside from this, some of the hypotheses from gst-kaldi-nnet2-online really tell me that the server seems to be corrupting the audio files.
Thank you very much!