This Device Teaches You to Avoid Filler Words

Filler words

Public speaking involves keeping an audience consistently engaged with what you are saying, and nothing breaks that level of concentration faster than using "filler" words such as "uh," "um," and "so." In order to train people to better avoid these words, Benedikt Groß, Maik Groß, and Thibault Durand worked together to create the "Mind the 'Uuh'" project, which aims to both count the number of times a filler word is spoken as well as alert the speaker. In doing so, the device hopes to train people to reduce their reliance on unnecessary words.

Components

Keyword detection requires ample amounts of processing power, which is why the team went with an Arduino Nano 33 BLE Sense. Even better, its onboard microphone negates the need for an external one to be connected. The counter is a standard four-digit, seven-segment display, and it is controlled by a 74HC595 shift register and a TPIC6B595 shift register at its drain. The counter is reset by pressing a red button, and alerts are sounded whenever the servo motor swings against a bell.

Training a model

After assembling the circuitry together, the team moved onto the firmware aspect of the project that would enable the detection of certain words. 1500 samples of the sound "uuh" were collected at a frequency of 16kHz and ranged between 300ms to 1s long in duration. Once collected within Edge Impulse Studio, they were run through an MFCC (Mel Frequency Cepstral Coefficient) filter, which isolates the frequencies humans communicate at. Lastly, the now-processed samples were used to train a Keras neural network and then exported as both an Arduino library and Web Assembly package.

The web application

Because the vast majority of people don't have access to the exact components the team selected for their project, they created a simple web application that uses the same model from the previous section but within a web browser. This means anyone with an internet-enabled device and a microphone can try saying "uuh" and get hear a little alert in response.

Seeing how it works

To test their new device, the team spoke several sentences with the word "uhh" repeatedly stated. As expected, the model was able to successfully determine when it was said, and this means other words could potentially be included in the future as well.

To see more about this project, you can view its GitHub repository here.