With the rapid development in ML and mobile devices, speech related technologies are booming.
Our team is interested in wake word detection which is a subfield of speech technologies:
WakeWord detection is the task of recognizing a predefined keyword for activating a speech assistant. In other words Alexa is always listening to you and it needs to be waked up to start talking.
There were three questions motivating this project:
1.How do we train a model detecting a wake word?
2.What are the qualities of a good wake word? Short vs Long? Hey vs Hi?
3.Can we train a model for any wake word, i.e. “Hey FourthBrain”?
In order to answer these questions:
*We created a wake word model engine:WakeME
*Provided methods and metrics for model eval
*Addressed questions on quality of good wake word
*Demonstrated the model capabilities with a web app.
Our study showed that:
*Longer words are less sensitive to model types
*Model type is more important than the length of the word
*There is a significant advantage of using "Hey" or "Hello" in the wake phrase
Steps to launch the WakeMe test app
Go into the wakemetestapp directory-
- cd ~./wakemetesapp
Init the app using the following-
- eb init -p docker-19.03.13-ce your-app-name --region us-west-2
Launch the app using the following-
- eb create your-app-name --instance_type t2.large --max-instances 1