IEEE patent Review

For this lab session we were divided into groups of five and we have to implement a practical application of the DSP.
Problem Definition:  Speech Recognition system. Individually we have to review one IEEE paper and one patent on the above topic.
Patent: US5715369
The patent selected talked about the testing system of speech recognition depending on different parameters such as voice gender, volume etc. A system having a host processor and a memory device where  memory contains audio files. The Vocabulary need to expand in response to previously known phrases and capturing the output from the said speech.

Paper: 
SPEECH RECOGNITION SYSTEM OF ARABIC DIGITS BASED ON A TELEPHONY ARABIC CORPUS

In  this paper automatic recognition of spoken digits is done. Spoken digit recognition is used in many applications such as airline reservation, telephone dialing. The Arabic language is different from the European languages. The pronunciation of the digits from zero to nine also differ. A hidden Markov model toolkit is used to implement the isolated word recognizer. 

Basic Operation of DSP Processor

A digital signal processor (DSP) is a specialized microprocessor, with its architecture optimized for the operational needs of digital signal processing. It is usually used to measure, filter or compress continuous real world analog signals. We used a custom board of C2000 processor. Code Composer studio was used as a coding platform. We also implemented the FFT algorithms. While implenenting on hardware we have to refer the registers which is not required in C language.

Frequency Sampling method for FIR filter design

Frequency Sampling method is also used to design a FIR filter.
In this method the the frequency response is sampled and the desired transfer function is obtained. The transfer function is a combination of FIR filter and parallel IIR filter. The final output is obtained using IDFT method. The final output is always symmetric about the point N/2. It is the most simplest method when desired frequency response is specified.

Windowing method for FIR filter Design

Applying a Window function helps us to get extra control on the characteristics of a filter. The desired filter is truncated with a window function. If the desired window is not given then it is decided depending on the values of As.
The different types of windowing functions are:
1) Hamming Window
2) Hanning Window
3) Rectangular Window
4) Blackman Window
5) Bartlett Window

Butterworth Filter

Butterworth filter is a type of signal processing filter designed to have a flat frequency response as possible in passband. 

The designing of butterworth filter was done using BLT(Bilinear Transform) method. Analog and digital filters were designed and frequency warping effect was studied.
The response of the filter has no ripple in stopband and passband. The poles are inside the unit circle and lie to the left hand side of s-plane. Thus both the filters are stable. For better stability the order of the filter should be increased.

Chebyshev Filter

Chebyshev filters are analog or digital filters and have more passband ripple.

Chebyshev filters are of two types:

1) Chebyshev-1: These filter are all pole filters in the passband, they have monotonic characteristics in the stopband

2) Chebyshev-2: This filter contains zeros as well as poles.

These filters minimize the error between the idealized and the actual filter characteristics over the range of filter but with the ripples in the passband.

OVERLAP ADD and OVERLAP SAVE methods

 Overlap-Add and Overlap-Save method is used to perform linear convolution and circular convolution of long data sequences respectively. They are called as block processing techniques as the input signal is divided into blocks and then analysed.  The output signal is the combination of the convolved input signals by removing first few values. The code is then executed by entering the input values and thus checking the result. 

FAST FOURIER TRANSFORM

 The Fast Fourier Transform(FFT) is used to reduce calculations and thus the process becomes fast. In this method the input sequence is broken down into two equal parts. This breaking down of the signal was continued till no further decomposition was possible. Radix-2 algorithm is used to perform the experiment. The counter was then added in the code to check the total number of calculations. Thus we conclude that FFT is computationally fast as the number of calculations are reduced.

DISCRETE FOURIER TRANSFORM

In the second lab we studied Discrete Fourier Transform(DFT) and its practical implementation. Three different cases were taken into consideration. In the first case DFT of four point signal is studied then in the second case DFT of zero padded signal is studied and in the last case DFT of expanded spectrum is studied and the magnitude of all the three is plotted. Thus we conclude that DFT produces periodic results. The spectrum is discrete and is in the range of [0,2π). As the length of signal increases due to zero padding the frequency spacing and approximation error decreases also the resolution of spectrum increases.   

CONVOLUTION and CORRELATION

The practical approach towards Convolution and Correlation is studied during the lab sessions. Convolution is a mathematical way of combining two signals {x(n) and h(n)}to form the third one{y(n)}.  Practicals on both linear and circular convolutions are performed. Circular convolution gives aliased output. In correlation we learned to find the degree of  similarities between the two signals. If the input is delayed the auto-correlation of delayed input is same as that of the original input signal. Thus all this is verified using different values if input signals.