DSPP

SINGLE CHANNEL SOUND SEPARATION Patent No.:  US 7,243,060 B2  Inventors: Les Atlas, Seattle, WA (US); Jeffrey Thompson, Bothell, WA (US) Publication Date: Jul. 10, 2007 Filing Date  : Apr. 2, 2003 The speech of two or more simultaneous speakers (or other simultaneous sounds) conveyed in a single channel are distinguished. Joint acoustic/modulation frequency analysis and display tools are used to localize and separate sonorant(a sound produced with the vocal cords) portions of multiple-speakers’ speech into distinct regions using invertible transform functions. For example, the regions representing one of the speakers are set to zero, and the inverted modified display maintains only the speech of the other speaker. A combined audio signal is manipulated using a base acoustic transform, followed by a second modulation transform, which separates the combined signals into distinguishable components. The components corresponding to the undesired sp...

Paper Title: Single-Channel Audio Sources Separation via Optimum Mask Filter Authors: Mahdi Fallah, Meysam Asgari In many single-channel audio separation techniques, expressing a given audio mixture in terms of its underlying signals is often introduced as a challenging topic. Opposed to commonly used GMM(Gaussian Mixture Model) based approaches, introducing a new mask filter for DFT mixtures, in this paper a VQ(vector quantization)-based single channel audio separation approach is proposed. Opposed to commonly used Mixture Maximization (MIXMAX) mask filter based approximation, which was extensively used in single-channel audio separation algorithm, it was demonstrated that the proposed approach outperformed MIXMAX mask filter in terms of SNR results as well as MOS results.

The DSP Processor on which this was implemented was TMS320F28375. Various instructions like 1. Arithmetic: ADDB, SUBB  2. Bitwise operations - AND, NOT 3. Shifting operation – LSL,LSR,ROR,ROL were learnt. Register values before and after execution of the program were noted.

This is the second method of FIR Filter design. The steps of finding the frequency response of the input is same as that of the windowing method. Then, DFT of the frequency response is calculated. The values are adjusted depending on the type of filter i.e. low pass filter or high pass filter. Inverse of the frequency sampled signal gives the desired filter design. The designing and plotting of magnitude response was done on Scilab.

A linear phase finite impulse response(FIR) filter was designed using Scilab. Filters like stop band and pass band attenuation and stop band and pass band frequency are taken as input from the user.There are different window functions that can be used like Hamming Window, Bartlet Window, Hanning Window, Blackman Window,etc. The window function used in this experiment is Hanning window.the formula is input in the Scilab code and then the program is executed.

We designed the Digital Chebyshev Filter using Scilab. Both Chebyshev LPF and HPF have been designed. We have plotted magnitude spectrum just like in case of Butterworth filter design. Ap, As, fp, fs and F have been entered and order of the filter and its transfer function were calculated. But in this case unlike the Butterworth filter plot, ripples were observed in the passband while the stop band was monotonic. The number of transitions in ripples signifies the order of the filter. It was also observed that the order of the filter was less in Chebyshev as compared to Butterworth for the same inputs. Hence Chebyshev filter requires less hardware. Originally posted at: https://chaitanya33.wordpress.com/2017/04/12/digital-chebyshev-filter-design/

We designed the Digital Butterworth Filter using Scilab. Both Butterworth LPF and HPF have been designed. The Passband Attenuation(Ap), Stopband Attenuation(As), Analog Passband frequency(fp), Analog Stopband frequency(fs) and Sampling frequency(F) were entered as inputs. The order and transfer function of the filter were calculated. The magnitude spectrum and pole zero plot were also observed. There were no ripples seen in the passband or in stopband. Butterworth filter has all its poles inside the unit circle. Hence the filter is stable. Originally posted at: https://chaitanya33.wordpress.com/2017/04/12/digital-butterworth-filter-design/

In a Digital FIR filter if the input sequence is long we use Overlap Add and Overlap Save Method. The input sequence is broken into shorter blocks and the output is computed block by block using convolution. In OAM since convolution is linear, the output of a long sequence can be calculated by summing the outputs of each block of the input. The tail of the output from the previous block is overlapped with the beginning of the output from the present block. In OSM the last part of the previous input block is used as the beginning of the next block and the overlapping portion of the output blocks are discarded. Originally posted at: https://chaitanya33.wordpress.com/2017/03/14/overlap-add-method-overlap-save-method/

An FFT produces exactly the same result as evaluating the DFT directly. The most important difference is that an FFT is much faster. A radix-2 decimation-in-time ( DIT ) FFT operates by decomposing an N point time domain signal into two N/2 point signal. This is repeated till we get 2 point time domain signals. In DITFFT the output sequence index is in order and the input sequence index is in bit reversed order of the output. In case of DIFFFT, it is the other way round. The number of calculations are reduced in FFT. Originally posted at: https://chaitanya33.wordpress.com/2017/03/13/fast-fourier-transform/

Discrete Fourier transform (DFT) takes a finite number of samples of a signal and transforms them into a finite number of frequency samples of that signal. We have used C programming and analysed three cases. A finite 4 point sequence was taken as input. The length of the output signal was same as input signal. When the signal is zero padded to an 8 point sequence, we observe that the resolution of the output magnitude spectrum improved. Now the input sequence is expanded to an 8 point sequence by adding alternate zeros. This compresses the spectrum in frequency domain. Also, DFT produces periodic results. Originally posted at: https://chaitanya33.wordpress.com/2017/03/13/discrete-fourier-transform/

Convolution is used to find the output of a system using input signal and impulse response. We have studied three cases: Linear convolution, Circular Convolution and Linear Convolution using Circular Convolution. The input signal x(n) and impulse response h(n) were entered to the computer program. The output signal y(n) was verified with the calculated values. It is observed that in LC, the length of output signal is N=L+M-1 where ‘L’ and ‘M’ are lengths of x(n) and h(n) respectively. While in CC, N=Max(L,M) and in case of LC using CC, N≥L+M-1. Correlation is used to find the degree of similarity between two signals. In case 1, we analyse auto-correlation of a signal. Here output signal y(n) is an even signal and y(0) is the maximum value. Auto-correlation of delayed input signal produced the same result as auto-correlation of original signal. Cross-correlation of input signal with delayed signal was observed to be same as advanced auto-correlated signal. The degree of similarity w...