Menu

PocketSphinx Accuracy

Help
Frank
2013-09-22
2015-09-15
  • Frank

    Frank - 2013-09-22

    Using PocketSphinx version 0.8 to convert a wav file to text I only get about a 50% accuracy.
    Running pocketsphinx_continuous.exe on a windows machine using the pre-built language files:
    -hmm model\hmm\en_US\hub4wsj_sc_8k
    -lm model\lm\en_US\hub4.5000.DMP
    -dict model\lm\en_US\cmu07a.dic

    Which model files are best to start working on in order to get better results?
    I did not see a way to attach the audio file so I just added the text.
    The file is a 16khz, 16 bit, mono recording.

    Here is the output result from Sphinx:

    for mayor it which are to cut phone calls for this year i've got don't have back i mean what we have seen however is that armed forces present a bomb and just the big about a week ago gave an interview at rolling stone where he said that he was planning to make climate change an issue in his campaign

    will see how it played out on clearly climate change does not why as the level of national priority for most americans as the economy and on employment and do you not to daughter kind of issues like i mean that's perfectly understand that wall and on the last because the parties have seen really terrific apart so far in terms of their views of climate changing climate change science that perhaps there's going to be over about discussion about what should be used to approach is to people prefer

    you know is a as this horse race gets going down but it down the stretch to this year word
    in a c. b. c. poll come out virtually every week rather have it and i'm asking you how many of them are men ask any of these these questions you will that's a ballot you know climate change global warming alternative energy that seems like to when you look at them calling the exit poll in what

    Here is what it should be:

    I mean, I wish I had a crystal ball for this year, but I don't, and a I mean what we have seen, however, is that, for instance, President Obama just think about a week ago gave an interview in Rolling Stone, where he said that he was planning to make climate change an issue in this campaign.

    We'll see how it plays out. um Clearly climate change does not uh rise to the level of national priority for most Americans. As the economy and unemployment and you know other kinds of issues like that mean that perfectly understandable but nonetheless because the parties have seemingly drifted apart so far in terms of their views of climate change in climate change science that perhaps there is going to be a robust discussion about which of these 2 approaches. Do people prefer.

    You know as a as this horse race gets going down the down the stretch here this year
    we're gonna be see be seeing polls come out virtually every week (laughter),
    but I'm asking you how many of them are gonna ask any of the these questions you asked about you know climate change global warming alternative energies. It seems like it when you look at the the polling the exit polling would.


    Any insight would be appreciated.
    Frank
    FrankE2u@hotmail.com

     
  • Nickolay V. Shmyrev

    You need to share an audio file

    It's better to use generic english LM for generic texts

     
    • Frank

      Frank - 2013-09-22

      Here is the attached audio file. It is 2mb. I can also email it.
      I will try the other english LM files.

       
  • Nickolay V. Shmyrev

    With

     pocketsphinx_continuous -infile Unemployment.wav -lm en-us.lm.dmp
    

    Result will be:

    it rail and richard your crystal ball for that your front pocket out and got i mean what we have seen however is that part is discredit all bomb adjusted to nick about a week ago gave an interview at rolling stone where she said that he was planning to make climate change an issue and campaign will see how it plays out on clearly climate change does not why's the level of national priority for most americans as the economy extend unemployment and to you the been the other type issues like i mean that's perfectly understandable but nonetheless because the parties have seemingly drifted apart so far in terms of their views of climate change climate change science that perhaps there's going to be robot discussion about which of these two approaches to people prefer still as that this horse race it's going down loaded down the stretch year this year word of the c. b. c. news poll come out virtually every week right after the i. i'm asking you how many of them are men ask any of the these questions you'll that's a valentino climate change global warming alternative energy to it seems like it when you look at them appalling of the exit polling what eh

    It's actually quite accurate and catched many words. For better results more advanced features could be used like speaker detection, online per-speaker adaptation. Accuracy should improve significantly on top of that but it's some work to setup the system properly.

     
    • Frank

      Frank - 2013-09-22

      Here is what I got when I used -lm en-us.lm.dmp -dict cmu07a.dic -hmm en-us
      Much better!


      and i will take care of all heard it here but i don't add back right we have however it bed out for an little bomb and get the money to buy a week ago gave an interview in rolling stone where he said that he was planning to make climate change an issue in campaign

      what the hell out laid out clearly climate change and not do ryan the level of national priority for most americans and the economy and on employment and the other kind of like i mean that's perfectly understandable but nonetheless because the party have drifted apart so far in an interview that climate change climate change climate that perhaps there's going to be about that and about which of these new approaches to people prefer

      it was a as this horse race gets going down and down the stretch year this year
      we're going to be see be seen polls come out virtually every week right but i'm asking you know how many of them are going as any of the these questions you asked about you know climate change global warming alternative energy is it seems like it when you look at them calling the exit polling what


      Thanks for you help.
      How do I implement the options you mentioned: speaker detectio, online per-speaker adaptation?
      Are there any command line options that would help?

       
  • Nickolay V. Shmyrev

    Are there any command line options that would help?

    There are no magic command line options.

    Feature integration requires coding to glue several packages.

     
    • Frank

      Frank - 2013-09-23

      So is there anything I can do using the standard tools to improve accuracy?
      For example, modify the acoustic model or dictionary?

       
  • Vincent N

    Vincent N - 2015-09-14

    Sorry to dig out this old thread.
    I took the wav file and ran it through pocketsphinx4 with the generic files and the output is really bad.

    the ad echoed off at here but i don't and dad and i knew what we added however it a god wanted a job ahmed and i think about a week ago gave an interview rolling stone where he said he was planning a big climate changes into his campaign for the hell up without palm oil including the novel lied to the level of natural priority for most americans at the economy and on employment and you that other competitors like diving that perfectly understandable but on the last because the parties have to three drifted apart all our interview their view of the climate the clinton iron that perhaps they're going to be a robot that got him about which indeed
    go to the people are still isn't as this horse race it's going down of the velma stretches your word of the c. b. seen calls come out virtually every week for half that what i'm asking you how many of them and ask any of the izzy these questions you asked about you know climate change global warming alternative energies it seems like it when you look at the appalling the exit polling would

    command is here:
    bin\Release\pocketsphinx_continuous.exe -infile Unemployment.wav -hmm model/en-us/en-us -lm model/en-us/en-us.lm.dmp -dict model/en-us/cmudict-en-us.dict > text-us.txt

     
  • Vincent N

    Vincent N - 2015-09-14

    switching back to the older hub4wsj_sc_8k much better:

    honey are richer have a crystal ball for this year part of guilt and back i mean what we have seen however is that our forces present all bomb adjusted to about a week ago gave an interview at rolling stone where she said that he was planning to make climate change an issue in this campaign will see how it plays out on clearly climate change does not know who was the level of national priority for most americans as the economy and on employment and you're not in the other kinds of issues like i mean that's perfectly understandable but nonetheless because the parties have even we drifted apart so far in terms of their views of climate change climate change science that perhaps there's going to be a robot discussion about which of these
    approaches to people prefer you know as as this horse race gets going down the new demonstrates year this year we're get receipt be seems hole come out virtually every week graphic know what i'm asking you how many of them are men ask any of the these questions us a ballot you know climate change global warming alternative energy it seems like it when you look at them polling in the exit polling what

    something wrong with the generic acoustic model ???

     
    • bic-user

      bic-user - 2015-09-15

      Well, the beginning of the file is a telephone speech. Looking at the spectrogram it seems to be 8 khz sampled. That's why you have poor accuracy with default 16khz model.

       
  • Vincent N

    Vincent N - 2015-09-15

    I am confused.
    When you take a 16000 Hz sampling rate file, of course it shows 8000 Hz on a spectogram.
    The sampling frequency is twice the audio bandwidth.

    When we say pocketsphinx need 16000 Hz it is the sampling rate, right ?

    So it should be fine unless I am mistaken by something else

     
    • bic-user

      bic-user - 2015-09-15

      No, the file is 16khz. But the first speaker in file is a telephone speech. It has gaps from 4 to 8khz, while 16khz acoustic model expects something to be there.

       

Log in to post a comment.