Tiny model for tiny system

Large model shows us the limitless perspective of what’s possible, but model doesn’t have to be big to do amazing things.

In this post, I am going to share a real example of even with similar level of resource usage as conventional method, better result can be achieved.

This noise suppression example model is fully deployed onto battery powered embedded device, all figures and recording are captured from live running device. Here is the setup.

Some details of the model:

Number of Parameters: 55k

Input & output format: a frame of 64 PCM data, 16kHz sample rate

Processing precision: int16

This is a raw PCM in PCM out model, no external processing required, making it plug-and-play. Following is deployed resource usage figures on two popular processor architectures.

Battery powered embedded device in products nowadays often has 512kB to 1MB of RAM, 100 to 300MHz processor. This model could comfortably fit in, along with other processing and application workload too. In my experience, this level of resource usage is on par with conventional threshold control adaptive filter approach, however, the listening experience is next level up.

Apart from doing what it says on the tin, removing a ton of noise, the NN approach offers couple unique advantages in the way of doing it:

• Processing is continuous and across the full signal bandwidth, preserving the desired signal to best extend.
• Immediate reaction, there is no control threshold to build up to, hence no fade-in fade-out distortion.

These result in a more natural listening experience, all within the same resource usage as the conventional method.

If you have the image in mind that NN model is resource monster, I hope this example help turning that around. However, deploying to tiny embedded systems has its unique challenges, it’s an area less looked after. Stay tuned and more of such topics will be covered in future posts.