Nuance ASR fine tuning

[sky_scrapper]

Hi i am using Nuance grammar to play a ten digit number.When i tell the number normally in voice,its accepting and sometimes its not accepting..I use fine tuning options like speedvsaccuracy(0.5),confidence level(0.2) in vxml and incompletetimeout in grammar...any other i am missing?I need to know the fine tuning options.

[msd86911]

-speedvsaccuracy:    balance between speed versus accuracy  (Set it to 1 to get best result but your server performance has to be good)
-sensitivity:              Specifies the level of sensitivity to speech (set it to zero in case o noisy environment)
-confidencelevel:      Specifies the speech recognition confidence level (0.2 is very low don't set it below 0.5)
-grammarmaxage:    Defines maximum acceptable age, in seconds, of cached grammar resources (Set it to 1s in case of development where you change your grammar a lot and want to cache it every time)

And you can always go to
<Nuance installation Directory>\Recognizer\data\callLogs\master\Resources\<IVR_Application_Name>\<Current_Year>\<Current_Month>\<Current_Day>\
it will hold collected audio file received by nuance and text file contain Recognizer result and its confidence even if it was rejected due to low confidence and didn't reach your MCP

[n3vek7]

I currently have the same kind of issues. I mean, we do ASR on a string of a letter and 14th digits, and digits are not recognized really good. Sometimes, 8 is inserted or 2 is duplicated.

- Audio recorded by Nuance sounds clean.
- Tried changing the option mentioned in this topic (speedvsaccuracy did help a little, but I still have the problem)
[code]<!-- Controls the trade-off between processing and accuracy. -->
  <param name="speedvsaccuracy">
    <declaration group="recognizer" type="float"
      set_by="default+api">
        <min_value>0.0</min_value>
        <max_value>1.0</max_value>
    </declaration>
    <value>1</value>
  </param>[/code]


[b]Here's an example : (I said C 1 2 3 4 5 1 2 3 4 5 1 2 3 4 )  [/b]
[code]TIME=20180305223605497|CHAN=FKJOAMIEAAANKPKAAAAAAAAA|EVNT=SWIrcnd|RSTT=ok|RENR=count|ENDR=ctimeout|NBST=1|RSLT=C81234512345123|RAWT=C 8 1 2 3 4 5 1 2 3 4 5 1 2 3|SPOK=C 8 1 2 3 4 5 1 2 3 4 5 1 2 3|GRMR=GURI0|KEYS=<SWI_confidence conf="0">0</SWI_confidence><__SWI_OUT_PROP__ conf="550">C81234512345123</__SWI_OUT_PROP__>|CONF=550|RAWS=-7561.079590|CMPT=false|BCNF=734|OCNF=140|SCNF=462|SPIV=2|SPAG=1|SPMS=en.us/10.0.2/models/SecondPass2/models2.hmm|WVNM=NUAN-35-32-SERVER001-FKJOAMIEAAANKPKAAAAAAAAA-utt002-POSTEP.wav|MDVR=482778560|NADP=0|CADP=0|LADP=N/A|MPNM=en.us/10.0.2/models/FirstPass/models.hmm|DPNM=NA|MACC=NULL|MEDIA=audio/basic;rate:8000|EOSS=10310|DURS=10310|EOSD=10360|BORT=6|EOST=9944|EORT=10211|EOFT=9945|CPRT=317|CPAR=0.100,0.056,1.343,0.577,1.655,0.094,0.509,1.000,0.000,1.669,-0.360|LA=idle|OFFS=0.000000|SCAL=1.000000|SRCH=SB:-75.0,WB:-78.0,PLB:-75.0,SPO:70.0,MA:10000|RCPU=343|UCPU=7015|SCPU=656[/code]

Any ideas? Fine tuning considerations?

Thanks !

[cavagnaro]

I have faced that issue here at Brazil because they do speak quite very different from region to region so I had to create my own grammar file instead of using the default one. Was tedious but worked fine.

[n3vek7]

Thanks cav,

Good point for the grammar, maybe it's too simple for now, we'll have a look at it again. In your case, were you using only numbers and letters?
I have some keyword, and there's no problem to recognize them (ex: Stop), but for the numbers, it's not as easy as expected.

I'll let know when we have a solution.

[cavagnaro]

Yeah, actually numbers are tricky as people tend to speak them in their own way. Or your input prompt has to be very specific on how you want them to input it, or do a grammar that covers all variations and possibilities. for example, credit card numbers was a real challenge