Menu

Pocketsphinx acoustic models

Help
Matigrab
2013-07-19
2013-08-19
  • Matigrab

    Matigrab - 2013-07-19

    Hi!

    I'm newbie in pocketsphinx and I don't know how to convert Voxforge acoustic models to pocketsphinx. I've tried to convert mixture_weights to sendump with python scipts (prune_mixw.py and quantize_mixw.py) but it doesn't work.

    Could anyone help me?

     
  • Nickolay V. Shmyrev

    I'm newbie in pocketsphinx and I don't know how to convert Voxforge acoustic models to pocketsphinx.

    You do not need any conversion, you can point to the model using -hmm option.

     
  • Matigrab

    Matigrab - 2013-07-22

    I'm trying to develop Android app based on PocketSphinxAndroidDemo example. Default used acoustic model is en_US. I want to change it to german acoustic model from voxforge.

    I've tried to simply past acoustic model files from voxforge-de-0.1\model_parameters, change -hmm, -lm and -dict paths and create language model and dictionary by lmtool but it doesn't work. The app stops working without any exception.

     

    Last edit: Matigrab 2013-07-22
  • Nickolay V. Shmyrev

    create language model and dictionary by lmtool but it doesn't work.

    Lmtool is for US English only, you need to use existing phonetic dictionary from the model. You have to create language model with language modelling toolkit yourself.

    The app stops working without any exception.

    Exceptions are in the file pocketsphinx.log

     
  • Matigrab

    Matigrab - 2013-07-22

    Exception from log file:

    INFO: acmod.c(246): Parsed model-specific feature parameters from /storage/sdcard0/CRModels/ac/vox_de/feat.params
    INFO: fe_interface.c(299): You are using the internal mechanism to generate the seed.
    FATAL_ERROR: "fe_sigproc.c", line 405: Failed to create filterbank, frequency range does not match. Sample rate 8000.000000, FFT size 512, lowerf 4765.625000 < freq -15.625000 > upperf 4218.750000.

     
  • Nickolay V. Shmyrev

    To record audio at 16khz you need to change 8000 to 16000 in sources in two places - decoder config option and AudioSource property

     
  • Matigrab

    Matigrab - 2013-07-22

    To record audio at 16khz you need to change 8000 to 16000 in sources in two places - decoder config option and AudioSource property

    Done! It's working properly.
    Now, I want to reduce dictionary and prepare language models for only few commands.

    Lmtool is for US English only, you need to use existing phonetic dictionary from the model. You have to create language model with language modelling toolkit yourself.

    There I found some tutorial. Should I build simply grammar? How to use voxforge model to prepare own language model and dictionary for few simply commands?

     

    Last edit: Matigrab 2013-07-22
  • Nickolay V. Shmyrev

    Should I build simply grammar?

    It depends on the complexity of the language you are going to recognize. For several simple commands you can write JSGF grammar and existing dictionary

    How to use voxforge model to prepare own language model and dictionary for few simply commands?

    Voxforge is an acoustic model, it is not related to language model

     
  • Matigrab

    Matigrab - 2013-08-19

    I would like to create simply language model and dictionary files with only few commands and use existing french acoustic model but it's too big for my mobile application.

    Is there any possibility to compress existing model?

     

Log in to post a comment.

Want the latest updates on software, tech news, and AI?
Get latest updates about software, tech news, and AI from SourceForge directly in your inbox once a month.