DSP Application

In this experiment, we had to perform an application on DSP processor. OT was a group experiment. The members were Chetan Jogi, Mohit Karandikar, Ameya Lokre, and Dhvanil Mandalia. Signal processing of a 1D signal was to be done. We chose the topic 'Audio feature extraction'. 

I reviewed the patent Audio processing techniques for semantic audio recognition.

Patent no.: US20140180673 A1

Summary:

A system or method was defined for determining the semantic info from audio. the incoming audio signl is sampled and processed to extract features like temporal, harmonic and rhytmic features. The extratced featurs are compared to stored reference signals. Features that are most similar to one or more templates from comparison are identified according to the tags. These tags are used to determine semantic audio data that includes genre, instrumentation, style, acoustical, dynamics and emotive description descriptor.

The paper I reviewed was Audio feature extraction for classification using relative transformation.

Summary:

Audio features that have been successfully used for audio classification include mel-frequency cepstral coefficeints (MFCCs), spectral similarity, timbral texture. Among these, MFCCs are widely used. Relative transformation has been applied to deal with sparse, noisy, imbalance problems, as it is learns the abilities from human beings. For each audio, a set of feature vectors can be constructed by computing MFCCs. One way of generating a single vector feature for the audio is to simply randomly subsmaple the overall distribution of these vectors. Here a new approach called Relative Transformation Vector Quantization (RTQV) is proposed.

We implemented a system which stores time domian reference signal and correlates it with a test signal. If the correlation output is 1 then the signal is identified.

https://drive.google.com/drive/folders/0BwvLoQVdPTK3SEJjVUxvbEd0bTg

Basic Operations Using DSP Processor

A BE student demonstrated his project on DSP Processor TMS320F28335. He programmed the processor using Assembly Level Language. After demonstrating his project, he showed how to perform various basic operations like Arithmetic, Logical and Shift operations. 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.

FIR Filter using Frequency Sampling Method

In frequency response of the obtained output signals of Freqeuncy Sampling Method, rippples in the Stop Band obtained are of decreasing amplitudes. Phase plot of HPF and LPF are same if the order of both of them are same. Phase varies linearly with freqeuncy , hence ouput will not be distracted.

https://drive.google.com/drive/folders/0BwvLoQVdPTK3OVBrbzFtYV9XeFE

FIR Filter using Windowing method

Depending upon the value of Pass band attenuation, appropriate window functin was chosen from Rectangulaer, Bartlett, Hanning, Hamming, and Blackmann. As stop band attenuation increases, main lobe width increases and side lobe width increases. Phase response varies linearly with frequency, thus no ditortin is observed at the output of the filter. Output is same as input delayed by some constant.

https://drive.google.com/drive/folders/0BwvLoQVdPTK3REp4cWtnN0VzdDg

Chebyshev Filter Design

Chebyshev filters have steeper roll-off and more passband/stopband ripples than Butterworth filters. Chebyshev filters minimize the error between idealized and actual filter characteristic over the range but with ripples in passband. There is a defnite ZERO at -1, number of poles gives the ordef of filter. Magnitude spectrum is equiripple in passband and monotonic in stop band.

https://drive.google.com/drive/folders/0BwvLoQVdPTK3eE85ZGFDWDJkN2c

Butterworth Filter Design

Butterworth filter is a type of signal processing filter designed to have as flat frequency response as possible in the passband. It is also referred to maximally flat magnitude filter. Poles lied in the unit circle for both HPF and LPF butterworth filters, thus they are stabe.

https://drive.google.com/drive/folders/0BwvLoQVdPTK3ZEp6OWphaUx6MjQ

Filtering of Data Sequence

The delay or latency period obtained while computing long datat sequences can be reduced by the two methods Overlap Add Method and Overlap Save Method. These methods dont require the enttire data sequence to be available before carrying ot the convolution.

https://drive.google.com/drive/folders/0BwvLoQVdPTK3Rk1JeUVaS2FWbG8

Fast Fourier Transform

Number of arithmetic calculations in Fast Fourier Transform is significantly less than that in Discrete Fourier Transform. Hence, Fast Fourier Transform is inevitably fasyer than Fast Fourier Transform.

https://drive.google.com/drive/folders/0BwvLoQVdPTK3Z1E4STd5MFJNUk0

Discrete Fourier Transform

As length of signal increases, frequency spacing reduces, approximation error representation of spectrum decreases, resolution of spectrum increases, missing values in less point DFT are present in more point DFT.

https://drive.google.com/drive/folders/0BwvLoQVdPTK3SzVXUW5CYWpvOWc

Convolution and Correlation

Length of output Convoluted signal equaled to sum of length of both input signals - 1. Length of Circular convoluted signal is the max of both input signals. Output of Auto correlated signal was same as that without delayed input signal. Length of output Cross correlation signal equaled to sum of length of both input signals - 1. Cross correlation with one input signal gives delayed output.

https://drive.google.com/drive/folders/0BwvLoQVdPTK3ZEVtTkJyY1ppeWs