synthesize audio -> Words not in lexicon about voice-builder HOT 3 CLOSED

So "Word not in lexicon" is a generic error we throw when something goes wrong , because 9/10 it is correct and persistent. So sometimes it might not actually mean "Word not in lexicon".

Note - if you are building a small voice, have at-least 50 sentences, anything below that is tricky.

Debugging

1 - Check whether there is a server called "model-server-55" under google compute engine instances

• If it isn't there, it means there was issue bring up a model server, under job 55 click "deploy", and see whether you can bring one up (stay for about 5mins to see whether there is a new instance called model-server-55),

  • If this doesn't bring a server up, it means something in GCE is preventing from bring up this server, can be quota, auth etc related. Easy way is to see the GCE notification and check whether anything there is a red. If not, goto GCP logging and under GCE check the latest error log.

• If the model server is launched, run a following curl, which should return "OK"

curl http://<ip>:3389/api/job/<job id>/model_server/healthz

If that doesn't work, it is a very weird network/firewall issue.

2 - False positive model - Model that was deployed to the model server is bad. Goto the job and open up the chrome network tab, synthesis some text, if response in the network tab for call - example - http://35.186.254.17/api/job/89/model_server/alignment?text=hello is like below, that means the model failed to train (debugging that is bit different)

exec error: Error: Command failed: ${FESTIVALDIR}/bin/festival -b festvox/goog_vb_unison_cg.scm  -b "(voice_goog_vb_unison_cg)"   -b "(Parameter.set 'Audio_Method 'Audio_Command)"   -b "(Parameter.set 'Audio_Command \"\")"   -b "(utt.relation.print (SayText \" asd \") 'Segment)"
Error reading ESPS file /home/model//festival/trees/goog_vb_unison_mcep.params
Cannot load track: /home/model//festival/trees/goog_vb_unison_mcep.params
SIOD ERROR: could not open file /home/model//festival/trees/goog_vb_unison_mcep.tree

from voice-builder.

Thank you for your help.

Good to know that the message "Words not in lexicon" is a generic error.

The "model-server-55" was not listed under my compute engine instances. There were no model-servers listed at all for the preceding jobs.

In the GCP notifications (all activities) the creation of the VM "model-server-55" and of the model servers related to most of my preceding jobs were however confirmed with two green messages "Create VM" and "Completed: Create VM" each.

In the "Compute Engine -> operations" list I found however red error messages stating for all jobs :

"operation: Create an instance"
"target : model-server-xx"
"status : SERVICE_ACCOUNT_ACCESS_DENIED"

I took a deeper look in the GCP documentation, mainly about IAM, service accounts, permissions and quotas, without finding hints about the problem.

Then I restarted a new project from scratch : "tts-lu-mb-2".

I checked and documented carefully every step done to deploy the new project as described in your "Voice Builder README".

SUCCESS

I received a new external IP address and created a Festival voice with the default Sinhala parameters and pathes. The related "model-server-1" is listed now in the "Compute Engine VM instances" and synthesizing words, copied from the lexicon, are played as expected.

curl http://35.240.4.177:3389/api/job/1/model_server/healthz sends an OK

I created afterwards two other voices, one with 50 prompts from the Sinhala data and one with 11 prompts from my luxembourgish data. Both jobs were deployed with success and synthesizing works. It's clear that the quality is poor because of the small size of prompts, but it validates the correct format of my files.

In the next steps I will focus to add more sentences to my luxembourgish data and regularly check the progress by creating new voices with the Voice Builder. I will keep you informed about my results.

Thank you again for providing this great tool.

from voice-builder.

Good to know it works now.
We have fixed the "Words not in lexicon" generic error, it should be more verbose now.
if you git pull origin master and then ./deploy.sh ui update you should have the update

Note - that the IPs are publicly accessible by anyone.

from voice-builder.