Voice PE wakes up - hears my question - then blinks indefinitely - fixed

Configuration Voice Assistant
,
1

Sharing an issue I found and was able to fix with voice PE, hopefully saves someone else a bit of time. The voice assistant PE, after initial configuration would hear the wake-up word, seemingly listen to my question then almost immediately two of the leds on it would start flashing until either reusing the wakeup word or pressing the button.

After determining the issue seemed to be coming from parsing the speech, I found logs at this endpoint: /config/app/core_speech-to-phrase/logs and stack traces indicating a words.txt file was improperly formatted:

ERROR: SymbolTable::ReadText: Bad number of columns (1), file = /share/speech-to-phrase/train/en_US-rhasspy/data/lang/words.txt, line = 194: econet
ERROR: FstHeader::Read: Bad FST header: standard input. Magic number not matched. Got: 0

Opening that file on HA under /share/speech-to-phrase/train/en_US-rhasspy/data/lang/words.txt reveals a map table of sorts that include a phrase and what looks like an index. Going to the line number that the stack trace complains about revealed a phrase that looked similar to this: econet^@^@^@^@^@^@^@^@^@ 193

Those ^@ characters apparently indicating null bytes. After reading a bit about what the file is, you can’t simply delete the offending line as the file is frequently regenerated. So, I traced this back to a sentences.yaml file under /share/speech-to-phrase/train/en_US-rhasspy/sentences.yaml and found that this value came from an area (like livingroom) defined in HA.

I went into the UI and updated the name of the area, making sure to highlight the entire value so that it removed any null characters. I saved the changes, and restarted the speech-to-phrase service. This regenerated the words.txt file leading to a bit more subtle version of the same problem. Stack traces indicating a different line number popped up, so I followed all the steps from above, except this time there were two identical values, which is not allowed:

valve 318
valve 319

In this case, the null characters never made it to the words.txt but were in the sentences.yaml. But also, since I had an area named econet and an econet valve, I put the pieces together and updated the valve name the same way I updated the area name. Restart speech-to-phrase service, and this thing finally works.

Sorry for the wall of text

2

Good catch

Speech-to-phrase trains on your named entities and areas at startup. Random characters and spaces often cause issues. Can be a bit of a pain to track it down but fix is simple.

3

Hi, welcome to the forum!

Your post becomes even more valuable if you mark it as solved.
You do that by selecting the three dots under the post:

image

Then select the check box:

image
By doing so:

  • this thread can be useful to other users as well (this thread comes up as solution when starting a thread which is similar to yours).
  • this prevents that someone else steps in trying to help

Thanks for giving back to the community! :+1:

4

appreciate that! I took a look and I’m thinking maybe the forum software has been updated or something because I don’t have an ellipsis on the main post:
ellipsis
I see a solution box on the replies, should I create a reply and accept that as the solution?

5

Hi, it makes sense that the OP cannot be tagged as the solution.
You could create a new post with how you solved it and tag that.

6

tl/dr solution here with longer walkthrough in OP:

Issue can be caused by null characters/whitespace in named entities that speech-to-phrase trains on. These can be addressed by updating the names in home assistant:

  • Check logs at <HOST>://config/app/core_speech-to-phrase/logs for errors in words.txt and get the line number it failed on
  • Review the line in that file on the system to see if it has weird characters or if the value is a duplicate somewhere else in the file
  • Trace to where to find these in HA by looking for the value in /share/speech-to-phrase/train/<language>-rhasspy/sentences.yaml and seeing what section it’s nested under
  • Navigate to the entity in HA and update the name making sure to select all in the edit field and renaming it.
    • can be the same name, just need to ensure any extra hidden characters are removed by highlighting the entire field (shift+end)
  • Restart the speech-to-phrase service
  • Rinse and repeat if necessary
2 Likes