Closed QAQ1551QAQ closed 5 years ago
seems to me that you train a model with 43 dim features (I think the aishell's mandarin speech contain extra pitch features), but you are using 40 dim mfcc as input
seems to me that you train a model with 43 dim features (I think the aishell's mandarin speech contain extra pitch features), but you are using 40 dim mfcc as input
So how should I modify my test input?
ah, it also might because of some issues in iVector
Cao you show the content of you online.conf here?
I think I met something similar like you, if there is a silence weight factor in the online.conf, I think you might have set it to 0, you can set it to a very small value like 0.00001.
I suspect the issue is that when you called prepare_online_decoding.sh you forgot to add the flags that relate to pitch features. You have to specify the pitch config somehow. Look at the script's usage message or the script itself.
On Fri, Aug 9, 2019 at 3:31 PM weiwchu notifications@github.com wrote:
I think I met something similar like you, if there is a silence weight factor in the online.conf, I think you might have set it to 0, you can set it to a very small value like 0.00001.
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/kaldi-asr/kaldi/issues/3511?email_source=notifications&email_token=AAZFLO6OZ5DWXZVZM2PYWETQDXA2FA5CNFSM4IKG7QP2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD37SKCI#issuecomment-520037641, or mute the thread https://github.com/notifications/unsubscribe-auth/AAZFLOYHX6T2GKD5ZEONFF3QDXA2FANCNFSM4IKG7QPQ .