Using CMU Sphinx acoustic model

[Chien Papier]

I have trained acoustic(and language) model(created by CMU Sphinx). How can I use it in Kaldi? Maybe there are some examples, I can't find them...Tutorial describes training from audios, but I haven't enough audios, only trained models.

[anand]

Why would you do this? "AFAIK" you cannot use CMU Sphinx's models with Kaldi. Even if you could, what do you plan to gain from using CMU Sphinx's models with Kaldi. 

[Chien Papier]

I haven't enough acoustic training data for my problem, I should recognize sentences from narrowly specialized area(mathematics). I tried to use CMU Sphinx's acoustic model and own language model(grammar) and it gave me good result. Now I want to build system for continuous speech recognition. But pocketsphinx recognizes continuous speech quite slowly. So, I decided to try Kaldi :)

But I haven't records of sentences I need. And I thought I could use CMU Sphinx's acoustic model somehow. I have only wav records of different stories(this is not what I need). Maybe I can use it for acoustic model training?(strongly doubt)

понедельник, 28 августа 2017 г., 12:35:07 UTC+3 пользователь anand написал:

[Wonkyum Lee]

I believe nobody can help you run kaldi with Sphinx's AM because it requires lots of engineering work and it does not give any gain from doing it.

If you don't have enough audio, just try librispeech or tedlium example it would be free for research purpose. But you have to train your-specific LM by yourself on top of AM. I think that's what you can do if you want to work with kaldi.

[Chien Papier]

I need Russian speech corpus, not English :) But thanks for the advice.

понедельник, 28 августа 2017 г., 19:48:02 UTC+3 пользователь Wonkyum Lee написал:

[Armando]

I believe Nickolay Shmyrev can help you for the Russian data