Increase Intent Recognition?

I am having an issue with Rhasspy understanding what comes after my wake word. The wake word is getting triggered properly, but then after I speak the command, it fails. The logs give me a no intent is recognized. Is there a way to increase Rhasspy’s understanding? I am using a Matrix voice, and it doesn’t seem to matter the distance. If I type the intent via the UI, it works fine

Can you give us more info on your setup?

  • profile (language)
  • Audio input mode (arecord, pyaudio, etc)
  • Kaldi or Pocketsphinx?

What is Rhasspy transcribing from your utterance (check the logs, it should be logged)?

Is there any error in your logs?

I am using english, and my wake word is the only thing that has been customized with snowboy (wake word doesn’t seem to be a problem). Everything else is default, including arecord and pocketsphinx. In the logs, it just says No Intent recognized. Here is a sample from the logs when i asked what time is it:

DEBUG:72912772] HermesMqtt: Published intent to hermes/nlu/intentNotRecognized
[INFO:72912782] DialogueManager: Automatically listening for wake word
[DEBUG:72912770] DialogueManager: handling -> ready
[DEBUG:72912732] WebSocketObserver: {“text”: “what five a”, “intent”: {“name”: “”, “confidence”: 0}, “entities”: [], “raw_text”: “what five a”, “speech_confidence”: 1.5322968200067263e-12, “wakeId”: “porcupine”, “siteId”: “default”, “slots”: {}}
[DEBUG:72912731] DialogueManager: recognizing -> handling

What hardware do you have? Looks like your voice is too low

Raspberry Pi 3B+ with a Matrix Voice on top, connected to an external speaker. Like my actual voice? For it to properly trigger the intent, I’m essentially yelling at the Rhasspy. Definitely not mumbling. And like I mentioned, its triggering the wake word fine, so that seems to indicate to me that the hardware is working appropriately (i.e. can hear and distinguish my voice)

Could you try to use kaldi?

    "speech_to_text": {
        "system": "kaldi"
    }
1 Like

Thanks! In my initial tests, kaldi seems to be working much better!

1 Like

So now I am getting an error message:

Graph not found at /profiles/en/kaldi/model/graph/HCLG.fst. Did you train your profile?

I was able to clear this the first time around, but it just popped back up. I looked through the forum, and it was recommended to put custom_word.txt in the kaldi folder, but that didn’t do anything. Any other ideas?

edit: after deleting the kaldi dir a few times and reinstalling, it finally created the HCLG file during training

1 Like