hi, following ur recipe ,result is good when training my data. but when i change feature into other tool's implement, like kaldi fbank and a torch implement librosa fbank(feature value is not same as librosa python implement), result is getting worse. is there any suggestion for a compareable result like librorsa python fbank implement. aslo i find the kaldi fbank looks not clear like librosa's, kaldi fbank is also successfully proved well in many audio task. but it's result getting worse when following ur recipe in vad task.
hi, following ur recipe ,result is good when training my data. but when i change feature into other tool's implement, like kaldi fbank and a torch implement librosa fbank(feature value is not same as librosa python implement), result is getting worse. is there any suggestion for a compareable result like librorsa python fbank implement. aslo i find the kaldi fbank looks not clear like librosa's, kaldi fbank is also successfully proved well in many audio task. but it's result getting worse when following ur recipe in vad task.