Thank you Andreas.

So here's the PRS slide.

I'm in the lab since February last year and I have one publication on my PhD topic which

I will be presenting today.

So when we look at our society and the people that are above the age of 85 we can see that

this highly correlates with the number of hearing disease patients.

So keeping that in mind, a good noise reduction and a good hearing aid keeps those patients

or enhances their daily life actually.

So when we talk about hearing aids and noise reduction we need to know a few constraints

that we need to consider.

So one is online processing.

So we can't use the whole signal as input for our algorithm.

We need to, we can't just wait one or maybe 10 seconds that we have the full sequence

to output a denoised or speech enhanced version.

And furthermore we have a very limited frequency resolution where multiple speech harmonics

could be in one frequency band and I will have an example very soon.

And obviously the power consumption and computational power is not that huge from hearing aids

but this is not yet considered in this work.

So the hearing aid pipeline that I'm working with is shown on this picture and we can see

in the beginning we have some kind of audio signal that is recorded with some microphones.

We transform it into frequency domain using an analysis filter bank and this is very similar

to a short time for ear transformation.

Then we do some directional processing which results then in a single channel signal and

this is used for our noise reduction and in the end the signal is transformed back to

time domain so we can listen to it using a synthesis filter bank.

So what is the actual input of our noise reduction algorithm?

Here we can see a high resolution spectrum of a speech signal and we can see the typical

speech harmonics from the voice parts of the speech.

But unfortunately our input looks like this where we don't see the speech harmonics because

our frequency resolution is very little and multiple speech harmonics can be within one

frequency band.

So how can we work with this?

To reduce the noise also within the different speech harmonics and to do that I want to

first introduce you to linear predictive coding.

So linear predictive coding provides a method to model harmonic and periodic signals and

here we see some synthesis signal and linear predictive coding then tries to predict the

next frame of the signal given a linear combination of the order n of the previous frames and

a coefficient a.

So this is just a linear combination and we can compute ideal coefficients a by minimizing

the error between the predicted signal and the real signal and solving an equation system

so we get a closed form solution for computing the coefficients a and then we can model the

signal quite good given only the last n frames and you might see it there are some the yellow

signals the reconstructed signal and the blue signal is the original signal and there are

some parts that are not reconstructed but a fully periodic signal can be almost perfectly

be reconstructed and when we look at the frequency response of the reconstruction we can see

that LPC models only the most relevant frequencies so for instance if you have some white noise

in there white noise is has the same magnitude for all frequencies so LPC should not model

those on this white noise.

How can this help us with a with our approach?

So the spectrogram that has a very low resolution also has some periodic structures so here