k2 generalization vs kaldi

[aliiire...@gmail.com]

Hi Dear,
In my experiments, E2E models don't have good generalization and they are very dependent on dataset but Kaldi models have good generalization however E2E models have better WER on in-domain datasets vs Kaldi.
In these cases, I prefer the Kaldi models. Kaldi models have a good variance of WER on in/out-domain datasets.
For example, I train a model with wav2vec, espnet, and Kaldi on a telephone dataset like swbd. WER of wa2vec on an in-domain dataset is 5% lower (Kaldi wer is 25% and wav2vec wer is 20%) but on a very easy clean out-domain dataset wav2vec wer is 25 % but Kaldi wer is 9%.

What is about K2-fst model generalization?

 
best regards

[Daniel Povey]

Hm, which specific model were you using in k2?

That's something we'll have to work on but I don't think it will improve super fast.

--
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/9b37820a-9b34-42b1-9fd4-da3cdf74bea3n%40googlegroups.com.

[Sage Khan (Sage Khan)]

Hi. Did you try out CNN-TDNN for Switchboard? 
Also check these experiments

Kaldi-based DNN Architectures for Speech Recognition in Romanian Alexandru-Lucian Georgescu, Horia Cucu, Corneliu Burileanu Speech and Dialogue Research Laboratory University Politehnica of Bucharest Bucharest, Romania

Georgescu et al. 2021, “Performance vs. Hardware Requirements in State-of-the-Art Automatic Speech Recognition.”

Cnn-tdnn-based architecture for speech recognition using grapheme models in bilingual czechslovak task in Text, Speech, and Dialogue, K. Ekˇstein, F. P´artl, and M. Konop´ık, Eds. Cham: Springer International Publishing, 2021, pp. 523–533. [12] J. V. Psutka, A. Praˇz´ak,

  

I played with this script of CNN-TDNN for my use on Code switched Urdu on telephonic and clean data mixed. https://github.com/anish9208/gramvaani_hindi_asr/blob/main/kaldi/asr/Run_cnn-tdnn.sh

[aliiire...@gmail.com]

Thanks,

I just want to start k2 training. what is the model you suggest?