DSPP Application on Audio Signal Separating

Patent Review : Signal Separating System

Link: https://patents.google.com/patent/US5835608A/en?q=audio&q=signal&q=separaitng

Summary:  The patent revolves around producing a system for separating signals such as intelligible sound, for example speech, from unwanted noise (background noise), such as random noise, when both of them are present in the same field. The system comprises of two transducers located within the field which both the sounds can pick up. Each transducer generates a representative output signal. The two transducers are in close proximity to each other. The intelligible sound is correlated. The noise is electronically attenuated.

IEEE Paper Review: An overview of informed audio signal separation

Link : http://ieeexplore.ieee.org/document/6616139/

Summary: Audio source separation consists in recovering different unknown signals called sources by filtering their observed mixtures. In music processing, most mixtures are stereophonic songs and the sources are the individual signals played by the instruments, e.g. bass, vocals, guitar, etc. Source separation is often achieved through a classical generalized Wiener filtering, which is controlled by parameters such as the power spectrograms and the spatial locations of the sources. 

For an efficient filtering, those parameters need to be available and their estimation is the main challenge faced by separation algorithms. In the blind scenario, only the mixtures are available and performance strongly depends on the mixtures considered. 

DSPP Processor Applications :

A new project was created and a simple code was written to demonstrate various arithmetic and logical instructions in DSP processor. Various operations like Arithmetic, Logical and Shift operations were performed.
The DSP kit used was TMS320F28375. The output of the instructions was stored in the registers and these values were obtained in real-time using the debugging functionality of the software. 
Values of registers were noted down before and after each execution. As the processor is made by Texas Instruments, the software used to program the processor is Code Composer Studio.

Design of FIR filter using Frequency Sampling Method

Frequency Sampling Method:

The aim of the experiment was to design FIR filter using Frequency sampling method. In this experiment, the various parameters like pass band attenuation, stop band attenuation, pass band frequency, stop band frequency and sampling frequency are passed as input and the order of the filter is calculated. The flow is Hd(w) to H(k) to h(n)

In this we convert the H(k) function into h(n) to plot the spectrum .In phase plot discontinuities were observed between lobes when plot goes out of the range of -pie to +pie.

Design of FIR filter using window method

The aim of the experiment was to design Linear Phase FIR filter using window function. In this experiment we used Hamming window function to design Low pass/high pass fiter. We plotted magnitude as well as phase response.

In this experiment the value of As and Ap are verified. As we go on increasing As depending on the filter the side lobe width decreases & main lobe width increases.

Chebyshev Filter Design

This experiment is performed on scilab. Scilab is one of the two major open-source alternatives for MATLAB.

By giving the input specification we plotted the magnitude plot for High pass and Low pass filter. From the magnitude spectrum it is observed that ripples are present in the pass band. Although it is monotonic is stopband.

In Chebyshev filter the number of ripples represent the order of the filter. 

Chebyshev filter has lower transition period as compared to Butterworth filter. Also, order of the Chebyshev filter is less than order of Butterworth filter. Therefore, less hardware is required for Chebyshev filter.

Butterworth Filter Design

Digital Butterworth Design:

In this experiment, we used Scilab for implementing the code. Scilab is one of the two major open-source alternatives for MATLAB.

In this experiment, the various parameters like pass band attenuation, stop band attenuation, pass band frequency, stop band frequency and sampling frequency are passed as input and the order of the filter is calculated. Similarly, the cutoff frequency is calculated.

The normalized transfer function is evaluated according to the filter type: LPF or HPF (replacing s by 1/s).

From the normalized transfer function, the denormalized function is calculated by substituting the value of cut off frequency. 

The frequency bands are complex.

Also, magnitude spectrum is monotonic i.e. varying in one direction.

Overlap Add Method & Overlap Save Method

For a larger valued sequence, we need to decompose the input signal as it is not possible to implement the data together at the same time. Hence, it is necessary to decompose the input signal into multiple finite signals to perform various operations. 

Thus for FFT, faster algorithms are used viz. Overlap add Method and Overlap Save Method.

In OAM, overlapping occurs but the overlapped portion is added to get the required output sequence. Whereas, in OSM overlapping occurs too but the overlapped portion is discarded. As we use convolution to find the output, decomposing helps in finding a more accurate output.

Also, OAM & OSM are equally fast methods to find the output of the FFT based input.

Fast Fourier Transform

FFT is a faster and efficient computational method for DFT. Fast Fourier transform reduces the number of computations required considerably when compared with DFT.

The signal is divided into 2 parts; the real and imaginary parts are calculated separately. The result is displayed as Z=X+IY i.e. Z= (real)+i(imaginary).
Thus, for a 4 point signal, two stages will be required which are executed using separate array multiplications.
FFT is used for parallel processing as it improves efficiency.

Discrete Fourier Transform

Discrete Fourier Transform is used to convert a Time domain signal into a Frequency domain signal. We used C programming language to implement the code.

We used arrays to store the values of coefficients of input signal. The real and imaginary parts of the output were computed independently and their values were stored in two different arrays.

By comparing the two DFTs (4 point and 8 point), it was found that the 8 point DFT was more accurate. So as N in N-point DFT increases the magnitude spectrum gets more closer to the actual spectrum which involves infinite length of x[n]. Thus, output of DFT is periodic.

Discrete Convolution & Correlation

The aim of this experiment was to study basic signal processing computations like Convolution and Correlation. Convolution is a mathematical way of combining two signals to form a third signal. We also need to know the basics of C programming.

The length of output of convolution y(n) was found to be one less than the length of the two input signals x(n) & h(n) by varying the length of one signal, keeping the other constant. 
Thus, the length N=L+M-1 (where N = Length of y(n), L = Length of x(n) & M = Length of h(n)). We also found that circular convolution gives aliased output i.e. the last few values get mixed (added) with the first.

For auto and cross correlation, we chose two signals; an input signal and its delayed input. Thus, correlation is used to fins similarity between two signals.

Also, zero padding is used in both convolution and correlation to get the required signal for processing.