Detect a crying child with tinyML and artificial information

When a child cries, it’s virtually at all times because of one thing that’s flawed, which might embody, amongst different issues, starvation, thirst, abdomen ache, or an excessive amount of noise. In his challenge, Nurgaliyev Shakhizat demonstrated how he was capable of leverage ChatGPT, text-to-audio, and tinyML applied sciences to construct a cry detection system with out the necessity for gathering real-world information himself.

The method is as follows: ChatGPT generates a collection of textual content prompts that every one contain a crying child in some method. These prompts are then handed to AudioLDM which creates sounds in response to the prompts. Lastly, Shakhizat used the Arduino Cloud’s Machine Studying Instruments integration, powered by Edge Impulse, to coach an ML mannequin for deployment onto a Nicla Voice board. To create the sounds themselves, Shakhizat configured a digital Python setting with the audioldm bundle put in. His script takes the listing of prompts, executes them inside an AudioLDM CLI command, and saves the generated sound information as a WAV file.

As soon as this course of was carried out, he configured a challenge within the Edge Impulse Studio which trains a classifier mannequin. The consequence after coaching accomplished was a mannequin that might precisely distinguish between background noise and a crying child 90% of the time, and deploying it onto the Nicla Voice confirmed the effectiveness of how artificial datasets and embedded fashions can be utilized in the true world.

To learn extra, you’ll be able to take a look at Shakhizat’s write-up right here on Hackster.io.