TDNN training

[pavitra kulkarni]

Hi all,

In aishell TDNN training, why pitch features are not given for i vector extraction?

--

Thanks & regards

pavitra kulkarni

[Daniel Povey]

it reduces latency to not use pitch, so it's easier to use in real-time applications; and if you use high-frequency-resolution features (like 80-dim MFCCs) the added benefit of pitch is not that great anyway.

--
Go to http://kaldi-asr.org/forums.html find out how to join
---
You received this message because you are subscribed to the Google Groups "kaldi-help" group.
To unsubscribe from this group and stop receiving emails from it, send an email to kaldi-help+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/CACU3-r-m1bHWPwVU79gC-jdxhJo8rC%3DR6cyEJLvwvV4StiY-bA%40mail.gmail.com.

[pavitra kulkarni]

Thanks for the reply.

Can i use MFCC+pitch features to train i vectors based speaker verification ?

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/CAEWAuyRAL2u0%3DZFX_-QR9PrNZFHCDH2gCg93%2BC9TnF47VVU-xA%40mail.gmail.com.

[Daniel Povey]

I don't recommend using pitch for speaker verification.

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/CACU3-r-%3DTTfu_z-dP0%3DghO-rn6qrs%2BHxAgZ6ur0GAHwCygAQDA%40mail.gmail.com.