I am not getting any text for decoding

[viju2008]

I have followed the steps given

However i always get the following output from the asr server

{"status":"ok","data":[{"confidence":0.862751,"text":""}],"interrupted":"endofspeech","time":1080}

Please guide on how to check the asr logs

[viju2008]

Sometimes i get only text as NO

[mikenewman1]

I think you might be seeing the same problem that I posted about in #31
If I switch in a model that I built in January, the recognition is great. With the latest Kaldi I get nothing but [NOISE] tokens
I posted a question to Kaldi-help
https://groups.google.com/forum/#!topic/kaldi-help/1N4aVb75IdU
but DP did not have any ideas

[mikenewman1]

I found the problem. In order to run with the latest (batchnorm) models you need to add a line after loading

    {
      bool binary;
      kaldi::Input ki(nnet3_rxfilename_, &binary);
      trans_model_->Read(ki.Stream(), binary);
      nnet_->Read(ki.Stream(), binary);

      // This is the crucial line
      SetBatchnormTestMode(true, &(nnet_->GetNnet()));
}

Note that this only affects newer models (built using Kaldi source from after about March 2017)
For full compatability with the latest Kaldi, these two are probably a good idea as well:

      SetDropoutTestMode(true, &(nnet_->GetNnet()));
      kaldi::nnet3::CollapseModel(kaldi::nnet3::CollapseModelConfig(), &(nnet_->GetNnet()));

This is shamelessly lifted from (eg) kaldi/src/online2bin/online2-wav-nnet3-latgen-faster.cc

[formigone]

I put some details on this same issue on #37 for what helped me get over this "issue."

[dpny518]

in which file do we add this line
SetBatchnormTestMode(true, &(nnet_->GetNnet()));

[mikenewman1]

In Nnet3LatgenFasterDecoder.cc

(in the function Nnet3LatgenFasterDecoder::Initialize)

[hc038]

@viju2008 I am in the same situation now. did you solve the problem?

[mikenewman1]

See the posts above. The code needed updating to support batchnorm. After this fix everything worked fine. Note however that I haven't used this code in years so it may be broken again.

[mikenewman1]

Sorry. You could try asking in the usual Kaldi help channel From: hc038 <notifications@github.com> Reply-To: dialogflow/asr-server <reply@reply.github.com> Date: Wednesday, November 11, 2020 at 6:49 AM To: dialogflow/asr-server <asr-server@noreply.github.com> Cc: "Mike Newman (SM)" <Mike.Newman@microsoft.com>, Mention <mention@noreply.github.com> Subject: Re: [dialogflow/asr-server] I am not getting any text for decoding (#32) @mikenewman1<https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fmikenewman1&data=04%7C01%7CMike.Newman%40microsoft.com%7C1bd19712d8fd4136732608d88637c887%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637406921437794104%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=haF058T30dz%2B%2BpZw%2B6zYcmzh%2FcL1NgTEDmlJgWPPvVc%3D&reserved=0> thanks for the quick reply, I have added that line to Nnet3LatgenFasterDecoder.cc but I am getting this error [Image removed by sender. image]<https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fuser-images.githubusercontent.com%2F73985177%2F98808303-e1850f00-2441-11eb-8876-38d61e92a838.png&data=04%7C01%7CMike.Newman%40microsoft.com%7C1bd19712d8fd4136732608d88637c887%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637406921437804103%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=wYwInNZGZ8AkKNZAiNFy%2FAS2PTst6jWBSRCdyHPD1l0%3D&reserved=0> — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub<https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fdialogflow%2Fasr-server%2Fissues%2F32%23issuecomment-725378875&data=04%7C01%7CMike.Newman%40microsoft.com%7C1bd19712d8fd4136732608d88637c887%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637406921437814099%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=H386stw5ruT5p9IStk5n7xDFbApqIniHGp5EJ6MBrts%3D&reserved=0>, or unsubscribe<https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fnotifications%2Funsubscribe-auth%2FADDS3TZTLTWARJVHVUK44V3SPJ223ANCNFSM4EIS5VYQ&data=04%7C01%7CMike.Newman%40microsoft.com%7C1bd19712d8fd4136732608d88637c887%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637406921437824091%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=4ktn1JzFYDD1%2Bhd9EaL2EYtkbo8eEntcL8RqjjcdoDM%3D&reserved=0>.

[hc038]

I am trying to do with the system mic(Recognition using web browser), does it automatically convert to 16000hz audio format?

[realill]

Javascript code downsamples browser input to 16000 https://github.com/dialogflow/asr-server/blob/master/asr-html/res/recorderWorker.js#L70

[hc038]

thanks Ilya.

[hc038]

This server is working fine with "curl" command but with "system mic(Recognition using web browser)" I only get this

any suggestions?

[realill]

• Ensure browser records data correctly, I believe there is a way to import recorded stream.
• Use Chrome Developer Console to debug javascript.
• I do not remember if javascript client uses multi-part to send data to server, but this maybe a difference between curl and javascript.
• You can emulate multi-part data sending with curl as well and see if it works.

By the end of the day if curl works you can write your own code to emulate what it does. But without multi-part you wont be able to productionize it very well. Multi-part allows to do "online" decoding where stream is decoded as you speak. So you better figure it out. ;)