• State TI Thinks Resolved
  • Locked Locked
  • Replies 4 replies
  • Answers 1 answer
  • Subscribers 103 subscribers
  • Views 689 views
  • Users 0 members are here
Support feedback
Options
Options
Similar topics

This thread has been locked.

If you have a related question, please click the "Ask a related question" button in the top right corner. The newly created question will be automatically linked to this question.

TMS320C5535: Training for New model with garbage: C5535

Abhi Giri
Prodigy 20 points
Part Number: TMS320C5535

Hi,

I am working on a speech recognition project in my college. For this I used C5535EZDSP board and tried out the demo application, but accuracy is not that great with that (Suggest any improvements if possible). In document I saw the models can be generated, and training can be done using TIESR. I went ahead and followed to generate both SmallVocabulary and Normal Vocabulary(phones). Below are words I am using:

open door

close door

lock

unlock car

I recorded 8khz speech audio using audacity and segmented same audio for both SmallVocabulary and Normal (phones). I have used 22 recording for each word.

Then after successful training, I ran both Vocabulary in C5535 hardware. It works great, but one problem it shows random commands even when nothing is said.

I figured this is due to absence of filler model. I saw few posts in TI forum related to that and tried to follow. But I am not able to do it.

In normal vocabulary, the _fil garbage doesnt work properly. Please suggest what is wrong? What parameters I should change?

In SmallVocabulary, I have put fill as garbage. Here it works 70-80% good. But sometimes shows "lock" as random, without me saying anything. Then if I blow air, itll show fill. Please suggest what I need to do here to fix this? Any parameters need to be changed in this?

I am fresher student in this. So will be glad if some can mentor me in this step by step process.

Thanks,

Giridhar

  • 0
    TI__Guru**** 161600 points
    Hi,

    I've notified the sw team. They will post their feedback directly here.

    Best Regards,
    Yordan
  • 0
    TI__Mastermind 30365 points
    Giridhar,

    The Tiesr demo isn't meant to be a high performance, highly accurate voice recognition engine. It is meant to demonstrate voice recognition on the DSP. So, please don't expect accuracy rates to be 100%. We typically ask customer to use Sensory's (3P) solution for more robust speech recognition.

    Having said that, please see
    e2e.ti.com/.../1952595

    Also see the demo in the CSL package which I believe has the _filler model
    C:\ti\c55_lp\c55_csl_3.08\demos\TIesr\c5535

    Lali
  • 0 in reply to Lalindra Jayatilleke
    Prodigy 70 points

    Dear Sirs,

    We just bought the TI dev.board and we are trying to do the same thing with changing the phrase words and improving the performance. We are checked sensory solution but their solution is too costly for us at this point. 

    We tried the same solution posted here but the performance didn't get better with the original phrase word "TI Voice Trigger". We wanted to know if there is a different version of CSL library that we should use instead of the latest version or what can we do to fix the problem. 

    e2e.ti.com/.../1952595

    Would you please help us how to do this step posted in above post?

     Step 3: Increased the codec gain configuration.

    We are very appreciated your support in advance. 

    Best Regards,

    M.S

  • 0 in reply to Mahsa Sasani
    TI__Mastermind 30365 points
    M.S,

    Have not looked into this extensively or played with the mic gain, but take a look at the file codec_aic3254.c and function CSL_Status AIC3254_init().

    This function has several gain parameters for the AIC.

    Lali