To predict both 8khz and 16khz data

[Farid Haziyev]

Hello and thank you for kaldi model and this amazing channel which makes life easier. My question is that I want my model to be able to predict both 8 khz and 16 khz audio files, should I train my model with 8 khz or 16khz audios ? 

[Daniel Povey]

I'd probably recommend 8kHz.

but 16kHz could potentially work too, as long as you train on mixed 8kHz (upsampled) and 16kHz.

I don't recommend to use ivectors in that case (or at least test the effect of them carefully).. I am concerned that some of the things we do with full-covariance Gaussians will

give you close-to-singular matrices when using upsampled data.

Dan

On Mon, Feb 15, 2021 at 6:45 PM Farid Haziyev <ferid....@gmail.com> wrote:

Hello and thank you for kaldi model and this amazing channel which makes life easier. My question is that I want my model to be able to predict both 8 khz and 16 khz audio files, should I train my model with 8 khz or 16khz audios ? 

--
Go to http://kaldi-asr.org/forums.html to find out how to join the kaldi-help group
---
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/142e76f3-a20c-4c37-bee6-a4612689bfd0n%40googlegroups.com.

[Daniel Povey]

... generally WERs will be better with a 16kHz model, if you have 16kHz data, which is why it may be worth considering a 16kHz model.