October 2021 Summaries

The dialog manager is a crucial component of modern conversational agents that determines the most fitting output given user inputs. Most implementations use speech-to-text functionality to convert audio inputs into text transcripts and produce text outputs for downstream processes. Voiceflow approaches this problem by allowing designers to specify an observable and explainable logical transition graph triggered by various user intents, simplifying the process through intent classification and entity extraction techniques. The Raspberry Pi conversational client uses Picovoice's wake word detection engine, Google Cloud Platform Speech-to-Text API, and a custom configuration file to integrate these components and create a functional conversational agent.

Creating an interaction model for an Alexa Skill involves setting up a natural language understanding (NLU) system, such as Microsoft LUIS, to recognize and interpret user inputs. To do this, one must create an Azure Cognitive Services account, which includes a Natural Language Understanding Service within Cognitive Services on the Azure Portal. The endpoint from this service is then used to interact with MS LUIS. Once the MS LUIS app is created, it can be trained and published to Staging, allowing for integration with Alexa Skills. A Lambda written Node.js code uses the @azure/cognitiveservices-luis-runtime package to manage requests from an OrderIntent handler, which retrieves values from a SearchQuery slot and sends them to MS LUIS for prediction. The result is managed by an intentDispatcher, completing the integration of external NLP into Alexa Skills.