Poor performance with nnet3 TDNN-F model

[LiamLonergan]

Hi all,

I'm developing an API for an Irish language ASR service, putting Kaldi models I've developed into production. However, I'm having difficulty reaching the sort of performance with Gstreamer as I get with the normal online2-tcp-nnet3... kaldi binary. Does Gstreamer do any downsampling that might be affecting how the audio is passed to the model? Nothing is jumping out as an obvious issue.

Any help or ideas would be greatly appreciated!