Dan Kaldi #5 Can we fine-tune ASR models in Kaldi by training it on more audio files?

Timecodes
0:15 fine-tuning means
0:50 have few recipes but performance is so so
1:50 recommend to train from scratch

Answer: NO
Recommended: Train from scratch using a mixture of the original training data and the in-domain data if that's possible.

Ask you questions here and we will try to ask Dan!
#Kaldı #DanielPovey #speechprocessing #stt #asr

[itaipee]

If the batch-norm is the main reason that acoustic fine-tuning does not work in Kaldi, then Kaldi model also should transcribe poorly out-of-domain audio ?
but this is not the case, from my experience, Kaldi models are pretty robust, relatively to their model-size at least