Intent Recognition

The Intent Recognition System is used to filter the intent of the spoken text by the user. It converts the transcript text from the Speech to Text system to a JSON with the intent of the user. In this project we are using Snips-NLU. It’s a system with very fast recognition speed and a moderate training time.

How to use Snips-NLU

Please install Snips-NLU with the instruction in Setup Software. In the service file from Snips-NLU you can set a path for the engine and dataset.yml and you there you can also set the language. We created our own install instruction, because the command rhasspy-snips-nlu is not installed in the Rhasspy virtual environment. So we decided to install it on our own and communicate with it via MQTT.

The following instruction does not work, it’s the old way to install Snips-NLU. To use Snips-NLU you need to add the following config to your profile:

"intent": {
  "system": "snips",
  "snips": {
    "language": "de",
    "engine_dir": "snips/engine",
    "dataset_file": "snips/dataset.yaml"
  }
}

You can setup de, en, es, fr, it, ja, ko, pt_br, pt_pt and zh as you language. If you don’t set the language, the profile language is used if you set one. The options engine_dir and dataset_file specifies where the engine directory, and the datasetfile are located in the profile folder.

Information sharing with Rhasspy

This information is also shared via MQTT. The STT publishes a recognition request on the topic hermes/nlu/query. When recognition is successful, the result will be published on the topic hermes/intent/<intentName>.
If the recognition is not successful, the answer will be published on hermes/nlu/intentNotRecognized. For more information please take a look on the MQTT API.