Experiences from the new modular setup about zamia-speech HOT 1 CLOSED

Hi svenha,

thanks for the feedback and the pull request, very cool! :)

The WER rates you got are very interesting and probably accurate. It simply show how big of an impact the LM really has. We went from ~3-4% WER to ~7% WER with the modular approach and my suspicion is that this is because we add each speech corpora sentence only once to the language model (instead of as many times as it occurs in the training material as we did before - and especially german voxforge prompts appear multiple times). I have not verified that, though.

Regarding the question what is "realistic" I am not sure for now. Given the impact the LM has it really boils down to adapting the model to your specific application's needs - which is what I am working on at the moment (see speech_kaldi_adapt.py if you're interested - lacks documentation and is not finished, though).

from zamia-speech.