Hi mohemara92,

couple of remarks:

• the audio recording is very distorted (you need to lower the recording volume)
• the audio recording is very long, try splitting it up into individual sentences
• we are measuring about 30-40% WER for the Sphinx PTM models for unknown speakers and background noise. Try the Kaldi models if you need better results
• try using a language model tailored to your application domain (that will help Sphinx as well as Kaldi)

cheers,

   guenter

Hi Guenter,

Thanks a lot for helping.

I have some questions related to your remarks:

• Is there an optimal volume level for the recordings? or it's dependet on the microphone and enviroment ..etc.
• How long should be one recording? because the attached one is only 30 seconds long.
• Are the ptm models dedicated to some known speakers? and if i need to recognize my voice, I have to train my own model.
• Regarding the Kaldi models, do you mean using the kaldi acoustic models with pocketsphinx or shift to another software? in case of the first, how can I find/build these models? I tried to search but found no kaldi acoustic models under https://github.com/uhh-lt/kaldi-tuda-de
  
  Kind regards,
  
  Mohamed
  

Hi Mohamed,

• Is there an optimal volume level for the recordings? or it's dependet on the microphone and enviroment ..etc.

as a rule of thumb when you check your recording in Audacity only few samples should peak outside the +- 0.5 range.

• How long should be one recording? because the attached one is only 30 seconds long.

30 seconds is kind of the upper limit but most of the training material is much shorter, i.e. in the 5-12s range (typically a single sentence).

• Are the ptm models dedicated to some known speakers? and if i need to recognize my voice, I have to train my own model.

actually I had built the PTM models on request (I am not using CMU Sphinx in any of my projects) so I have very little experience with them. From what I see in the training results it seems to me that you cannot use them as-is for general unknown speaker speech recognition. You will either have to adapt them to a specific voice or tailor the language model for a narrow domain to get useful results (or, even better: do both ;) )

• Regarding the Kaldi models, do you mean using the kaldi acoustic models with pocketsphinx or shift to another software? 

you will have to switch to kaldi-asr to use those

• in case of the first, how can I find/build these models? I tried to search but found no kaldi acoustic models under https://github.com/uhh-lt/kaldi-tuda-de