DSP Appication: IEEE Paper and Patent reviwe

For this lab session, we were divided into groups and had to implement a practical application of signal processing.

Team : Parth Panchal,Soumyaa Passari ,Pranali Patil,Purvika Patil and Suma Salian
Individual experiment was to review one IEEE paper and one patent on Speech Recognition.

IEEE Paper:Hardware Implementation of Real-Time Speech Recognition System using TMS320C6713 DSP J.Manikandan, B.Venkataramani, K.Girish, H.Karthic and V.Siddharth Department of Electronics and Communication Engineering (ECE), National Institute of Technology, Trichy (NITT), Trichy-620015, Tamil Nadu, INDIA 

Year of Publishing: 2011

Where:  24th Internatioal Conference on VLSI Design

Review:

In this paper, hardware implementation of real-time speech recognition system on a Texas Instruments TMS320C6713 digital signal processor is proposed. Two approaches, one using MFCC feature and another using ZC feature from Cochlear are adopted.

Patent review

Application no: US 07/931,240

Patent no: US5353377 A

Date of Patent: Oct 4, 1994

Inventor: Akihiro Kuroda, Masafumi Nishimura, Koichi Toshioka

Summary: 

This invention relates to speech recognition and, more particularly, to a system capable of performing real time recognition of a large vocabulary by means of a signal processing card packaged on a bus of a personal computer.

Learning Experience: Operations on DSP Processor

In this experiment we studied assembly level programming using TMS320F2835 kit. We performed arithmetic,logical and shift operations. The changes in registers before and after operations were observed. We used code composer studio for this.

Learning Experience: FIR Filter Design Using Frequency Sampling Method

In this experiment, we were required to design a digital FIR filter using Frequency Sampling Method (FSM). Scilab was used to implement it.Thus, for the same values of attenuation in stop band and pass band, pass band and stop band frequencies and sampling frequencies we observe that the order of the FIR filter is much higher compared to that of the IIR filter.Phase spectrum is linear within positive lobes of magnitude spectrum.It is discontinuous in hase spectrum between lobes and also when phase spectrum goes out of range.

Learning Experience: Filter Design with Window function

The user had to input values like Attenuation in Stop band (As) and Pass band (Ap) as well as Pass band frequency, Stop band frequency and sampling frequency.The window function is required to truncate infinite samples of hd(n). Different window functions are Rectangular, Bartlett, Hanning, Hamming and Blackman with Blackman having the largest As and rectangular the least. Window method is widely used because different window functions can be used for different specifications of filter design A high pass filter and a bandpass filter was designed.The phase  plot being linear there will be no distortion at the output.Thus it seen that with increasing As the accuracy increases

Learning Experience: Chebyshev FIlter design

Low pass and High Pass filters were designed with the input parameters As ,Ap pass band frequency and stop band frequency and sampling frequency.The order of the filter was calculated using the given values of attenuation in pass band and stop band, pass band and stop band frequencies and sampling frequency. The normalised and denormalised H(s) was calculated, from which the Transfer Function H(z) was calculated.
From the magnitude response of digital Chebyshev filter, we observe that the number of peaks and valleys in pass band equals the order of the filter.

It was also observed that for the same input specifications for Butterworth and Chebyshev Filter, Chebyshev filter had a smaller order, which means it is better for real world applications as less order translates to less components required to build the filter and, hence, lest cost.  In chebyshev filter, the poles are plotted on an ellipse with imaginary axis as major axis in the s-plane.

Learning Experience: Butterworth Filter Design

Low pass and High Pass filters were designed with the input parameters As ,Ap pass band frequency and stop band frequency and samppling frequency.The pole zero plot was also drawn and it was observed that all the poles lie inside the unit circle indication both the High Pass and Low Pass fiters were stable.It was observed that  the butterworth filter is monotonic in it's pass band as well as stop band meaning it does not have any ripples.

The final steps are calculating the normalized and denormalized transfer functions followed by finding the z transform of the denormalized transfer function.

Our first experience of designing filters and from the output of both the filters was that the waveform of the filters are flat and ripple-free.With higher order, the magnitude response became sharoer and started resembling the ideal filter more closely

Learning Experience: Overlap Add Method and Overlap Save Method

Experiment Performed and Conclusions Inferred:

We performed block processing using two methods, overlap add method and overlap save method. 

Overlap-Add method: In this method, the signal was decomposed into smaller signals depending upon the number of samples that allows circular convolution to be equivalent to linear convolution and zero padding was done prior to the FFT to prevent the wrapping around of values. Using Fast Fourier Transform method, linear convolution of the input signal was found. The output of the signals result into overlapping of the values for the same indexes, hence the final output consists of the addition of all of the overlapped outputs.

Overlap-Save Method: In overlap-save method the last part of the previous block is used as the first part of the next block. No zero padding is done of the data, hence the circular convolution is not equal to the linear convolution. The output of the convolution of different blocks are wrapped around. These portions are useless and they are simply discarded.  

While the computational time for OAS and OAM is the same, since OSM does not require the addition of the overlapped output, it is faster than OAM.