Deep Neural Networks (DNN) have been widely used in speech processing, where benefits in speech recognition have been studied. Improvements in noise suppression using neural networks show great potential, but a large number of parameters, and especially the (weight) of the network makes it difficult to implement when low resources are available. This is the case when a Neural Network is implemented on a smartphone where the memory is fixed and processing time needs to be as fast as possible. The advances in deep learning are the starting point of the following three projects: DNN Compression, Wind Noise Suppression (WNS), DNN for acoustic echo cancellation.
CLIENT PROFILE
The world leader in secure connectivity solutions for embedded applications, Our client is a driving innovation force in the automotive, industrial & IoT, mobile, and communication infrastructure markets.
CHALLENGES
DNN compression
The target of this project is to investigate how to compress the Neural Network, reducing the number of parameters to fit in less than 500kB.
Implementation of different techniques needed to be done in python using state of the art algorithms from similar fields.
Wind Noise Suppression
During this project – that is actually part of a bigger plan with more people involved in the team – we are examining different solutions for voice recordings. As you might have experienced: recording videos outside can make voice/music less clear due to the noise coming from the wind for example.
If we can reduce this noise, we create a better user experience. Therefore the challenge contained the extension of the actual solution from 2 microphones up to 3 and 4 in order to match the client specifications.
DNN for acoustic echo cancellation
Given the benefit seen in noise suppression with NN, we explored the possible improvements that can be achieved using NN for echo cancellation. This time, our goal was to obtain an end-to-end model to suppress the echo coming from the far-end speaker.