Menu

Error in training

Help
2020-01-14
2020-01-14
  • Nikita Kumari

    Nikita Kumari - 2020-01-14

    Hello,
    1) I have trained CMU sphinx on my own model, sentence error rate and word rate are low but at the time of recognition text is not accurate because i'm getting error throughout training.
    2) While training i'm getting error messgaes to see log directoy, there i'm not able to get what has gone wrong.

    Baum welch starting for 1 Gaussian(s), iteration: 1 (1 of 1)

    bw Log File

    ERROR: This step had 8 ERROR messages and 0 WARNING messages. Please check the log file for details.
    completed

    Normalization for iteration: 1

    norm Log File
    completed

    Current Overall Likelihood Per Frame = -151.338177689961

    Baum welch starting for 1 Gaussian(s), iteration: 2 (1 of 1)

    bw Log File

    ERROR: This step had 8 ERROR messages and 0 WARNING messages. Please check the log file for details.
    completed

    Normalization for iteration: 2

    norm Log File
    completed

    Current Overall Likelihood Per Frame = -151.313717588741

    Split Gaussians, increase by 1

    inc_comp Log File
    completed

    Current Overall Likelihood Per Frame = -151.313717588741

    Convergence Ratio = 0.0244601012200008

    Baum welch starting for 2 Gaussian(s), iteration: 1 (1 of 1)

    bw Log File

    ERROR: This step had 8 ERROR messages and 0 WARNING messages. Please check the log file for details.
    completed

    Normalization for iteration: 1

    norm Log File
    completed

    Current Overall Likelihood Per Frame = -151.746004922352

    Baum welch starting for 2 Gaussian(s), iteration: 2 (1 of 1)

    bw Log File

    ERROR: This step had 8 ERROR messages and 0 WARNING messages. Please check the log file for details.
    completed

    Normalization for iteration: 2

    norm Log File
    completed

    Current Overall Likelihood Per Frame = -151.201296450361

    Convergence Ratio = 0.544708471991498

    Baum welch starting for 2 Gaussian(s), iteration: 3 (1 of 1)

    3) In reference to previous threads, i checked properties of my .wav files

    General
    Complete name : hi2505M0145.wav
    Format : Wave
    File size : 219 KiB
    Duration : 7s 0ms
    Overall bit rate mode : Constant
    Overall bit rate : 256 Kbps

    Audio
    Format : PCM
    Format settings, Endianness : Little
    Format settings, Sign : Signed
    Codec ID : 1
    Duration : 7s 0ms
    Bit rate mode : Constant
    Bit rate : 256 Kbps
    Channel(s) : 1 channel
    Sampling rate : 16.0 KHz
    Bit depth : 16 bits
    Stream size : 219 KiB (100%)

    please help me out in figuring where the problem is?

     
    • Nickolay V. Shmyrev

      The error

      ERROR: “backward.c”, line 430: Failed to align audio to transcript: final state of the search is not reached.

      is covered in troubleshooting section of the tutorial

      https://cmusphinx.github.io/wiki/tutorialam/#troubleshooting

       
  • Nikita Kumari

    Nikita Kumari - 2020-01-14

    Sir, I read this section of the tutorial. But can you please provide the steps to recover this?
    I mean how to do training with forced alignment?
    In sphinxtrain.cfg, i have changed forced_alignment = yes but we need forced_alignment dictionary also, how to prepare this?

     
    • Nickolay V. Shmyrev

      1) listen the file
      2) read the transcript and try to find the mismatch in the transdcript
      3) fix the transcript in transcription file
      4) start the training again.

      You can also ignore this error, tutorial says it is not critical.

       
  • Nikita Kumari

    Nikita Kumari - 2020-01-14

    i can ignore forced align as my sentence rate and word error rate is coming out to be good but when i'm using trained model recognition is not accurate.

    sentence error 26.4% and WER 4.4%

     
  • Nikita Kumari

    Nikita Kumari - 2020-01-15

    Thank you so much for your help. I corrected my transcript and it worked!

     

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.