Final Year Project Progress Report

Speaker Recognition/Verification Software for use in Security Applications Matthew Byrne 05362725 Supervisor: Dr Edward Jones Co-Supervisor: Dr Martin Glavin

FYI Progress report

Contents
1. INTRODUCTION 1.1 Project Background 1.1.1 Applications of Speaker Recognition/Verification software 1.1.2 Theory behind Speaker Recognition/Verification software 1.1.3 Project Aim 2. Technology Background 2.1 Matlab 3. Progress to date 3.1 Front-end Processing 3.2 Mel-Frequency Cepstral Coefficients 3.3 Input 4. Remaining Project Goals 4.1 Testing 4.2 Investigation of Speaker Recognition and verification over the Internet 4.3 Investigation and Development of real-time implementation 4.4 PC GUI APPENDIX I CODE 1. Program code 2. MFCC code APPENDIX II - TABLE OF REFERENCES

12 Jan. 09

Matthew Byrne

FYI Progress report

Chapter 1 Introduction
1.1 Project Background
1.1.1 Applications of Speaker Recognition/Verification software
Speaker Verification software, or voice biometrics, is technology that relies on that fact that we all have a unique voice that identifies us. Speaker verification software is most commonly used in security applications, such as gaining access to restricted areas of a building, or computer files by using your voice. It is also used in call centres where clients say the menu which they wish to select. In many applications, speech may be the main or only means of transferring information and so would provide a simpler method for authentication. The telephone system provides a familiar network for obtaining and delivering the speech signal. For telephony based applications, there would be no need for any equipment or networks to be installed as the telephone and mobile network would suffice. For non-telephone applications, soundcards and microphones are available as a cheap alternative.

1.1.2 Theory behind Speaker Recognition/Verification software

There are two different methods in speaker Recognition: - Verification: Determining whether an unknown voice is from a particular enrolled speaker. - Identification: Associating an unknown voice with one from a set of enrolled speakers. Speaker verification systems can be broken down into either text-dependant or textindependent systems. Most systems are based on scenarios where users repeat prompted preset phrases or speak a fixed PIN. These employ what is know as a text-

12 Jan. 09

Matthew Byrne

FYI Progress report

dependant system. Text-dependant systems can greatly improve accuracy, though such constraints can be hard or impossible to enforce in certain cases. Text-independent systems analyse and uses any spoken word to identify the user. A speaker verification system comprises of two states, a training state, and a test/verification state: The first step of training is to transform the speech signal to a set of vectors, and to obtain a new representation of the speech which is more suitable for statistical modelling. Firstly the speech signal is broken up into short frames, then windowed to minimize distortion. The signal is then analysed and stored as that users template. The first step in the testing stage is to extract features from the input speech, similar to that during training, compare the input speech to all other stored templates and select the most accurately matching template and ID the speaker. The figure below charts this process.

12 Jan. 09

Matthew Byrne

FYI Progress report

1.1.3 Project Aim

The main aim of this project was to investigate and build a speaker recognition system that recognises a person based on his/her speech patterns, capable of use in security applications. Such a system could be used, for example, in a security application where the identity of a user is determined based on analysis of a pre-determined (text dependent) or unknown (text independent) phrase. Also, speaker recognition systems differ slightly from speaker verification systems, in particular recognition involves choice of an identity from a set of possible users, while verification involve confirmation of a claimed identity (i.e. a yes/no answer). Such systems typically involve two main components: a Front End Processor (FEP) which processes the spoken input and extracts useful features from the signal, and a Classifier, whose function is to examine the features and make a decision based as to the identity of the speaker. Various approaches exist for the FEP, most of them based in some way on spectral analysis of the speech signal. Examples include Mel-Frequency Cepstral Coefficients (MFCCs), and Linear Prediction Cepstral Coefficients (LPCCs), and variations thereof. Likewise, there are a number of possible approaches for the Classifier including Artificial Neural Networks (ANNs), as well as various statistical methods.

12 Jan. 09

Matthew Byrne

FYI Progress report

Chapter 2 Technology Background

2.1 Matlab
Matlab is a numerical computing programming language. It allows easy matrix manipulation, plotting of functions and data, implementation of algorithm, creation of UI, and interfacing with programs in other languages. It allows for other toolboxes to be included for additional capabilities. Typical uses of Matlab include: Math and computation Algorithm development Modelling, simulation, and prototyping Data analysis, exploration, and visualization Scientific and engineering graphics Application development, including Graphical User Interface building

12 Jan. 09

Matthew Byrne

FYI Progress report

Chapter 3 Progress to date

3.1 Front End Processing
The front-end processing stage is performed in Matlab. Front-end processing is vital as it off-loads the computational data from the rest of the system. The front-end processor is used to extract useful details about the speech data from the input audio file (in this case it extracts MFCCs). The speech signal is passed into the FEP, and Matlab reads the input speech from a binary format file, or other audio file formats depending on what functions are used in Matlab to open such files. Next, the Frame length, sampling frequency, and number of channels which are required are passed into a function to compute a matrix of mel filters which will be used with the input signal to analyse the data. The characteristics of speech are relatively stationary over short periods of time, over longer periods of time an overflow of speech input makes extracting data more difficult. The input signal is then split into approximately 30ms frame lengths.
2000

1500

1000

500

-500

-1000

-1500

50

100

150

200

250

300

Figure: ~30ms Frame

12 Jan. 09

Matthew Byrne

FYI Progress report

The next step is to apply Hamming windowing to minimize the signal discontinuities at the beginning and end of each frame. This helps minimize the spectral distortion by using the window to taper the signal to zero at the beginning and the end of each frame.

800 600 400 200 0 -200 -400 -600 -800 -1000

50

100

150

200

250

300

Figure: Signal with Hamming windowing The signal is then converted to the frequency domain using a Fast Fourier transform. The first 128 frames are taken and plotted against the MFCC filter bank.

12 Jan. 09

Matthew Byrne

FYI Progress report

2 1.8 1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0

x 10

10

12

14

16

18

20

Figure: FFT of Signal analysed against MFCC

12 Jan. 09

Matthew Byrne

FYI Progress report

3.2 Mel Frequency Cepstral Coefficients (MFCCs)

In sound processing, MFCCs are based on the known variation of the human ears critical bandwidths with frequency, filters spaced linearly at low frequencies and logarithmically at high frequencies are used to capture phonetically important characteristics of speech. The Mel Frequency Cepstrum spaces the frequency bands equally on the mel scale, which approximates the humam auditory systems response more closely then the linearly-spaced frequency bands used in the normal cepstrum. Each filter in the MFCC has a magnitude response which is a triangle covering a certain range. The triangular is used to weigh a piece of the spectrum, and then the weighted values are summed together to give the overall filter output. The start of each filter is equal to the centre frequency of the preceding one, and the end of each filter is the
0.5

centre of the following one (the first filter starts at 0 and the last one ends at 4000). Suppose youre dealing with the 12th, filter, which has a centre frequency of 1292 Hz. The lower slope of the triangle goes from 0 to 1 between 1137 and 1292 Hz, while the upper slope of the triangle goes from 1 back
0.15 0.1 0.05 0 0.45 0.4 0.35 0.3 0.25 0.2

10

12

14

16

18

20

down to 0 between 1292 and 1469 Hz.

12 Jan. 09

Matthew Byrne

10

FYI Progress report

3.3 Input
The project will initially involve algorithm development and simulation using Matlab, and using pre-existing publicly available databases of speech. However, the project will also involve the development of a system for real-time operation, using a commerciallyavailable embedded microprocessor development system or a system based on a dedicated Digital Signal Processor (depending on the choice of hardware, only a subset of the functionality of the full system may be implemented). Initially a text-dependent, speaker identification system would be developed. Input speech data is required for both training and testing. The speech used for initial training and testing was pre-recorded. The system initially requires the pre-recorded speech file. In latter stages, the system will use recorded audio from multiple users. This will require a microphone for input which will store the signal temporarily while the system analyses it against stored users. In the final stages of the project, the system will wait for a prompt to start recording. The system will then wait for audio silence to stop analyzing. This audio silence will then be removed from the signal and then be analysis as usual.

12 Jan. 09

Matthew Byrne

11

FYI Progress report

Chapter 4 Remaining Project Goals

4.1 Testing
The current matlab code will be tested to see how robust it is at speech recognition.

4.2 Investigation of Speaker Recognition and verification over the Internet

With telecommunications inevitably moving towards an ip address based system, speaker recognition and verification over the internet will soon become a bigger factor in day to day life. VOIP uses UDP instead of the usual TCP for the following reasons: TCP is too heavy for real-time applications. TCP requires a handshake and a session between sender and receiver. UDPs task is to deliver data as quickly as possible, and has no control over the order in which the packets arrive at the destination. It is the receivers job to sort the order in which the packets rearrange. If packets take too long to transmit, or get lost in the transmission the receiver will ignore them and wait for the next packet to arrive. In voice transmission it is better to lose a few packets, rather then to wait/hang for those packets to be received. Companys such as blueface offer VOIP through our home phones, utilizing the preexisting telephone system, making implementing a speaker recognition and verification system simpler.

4.3 Investigation and development of real-time implementation

This includes selection of a suitable development platform based on performance requirements, ease of development etc. Translating the program from matlab code to C code would provide a more universal language and hence more flexibility in the speaker recognition and verification software. Finally the development of a real-time version of the system, and testing/evaluation of the system in a laboratory environment shall be carried out.

12 Jan. 09

Matthew Byrne

12

FYI Progress report

4.4 PC GUI
A program with a graphical user interface will be coded. Users will have the options to train or run the program. Training will require a login from a previously trained user with a sufficient security profile. The training protocol will then be run. When testing is selected the user will be prompted to speak a given passphrase, and based on the analysis the user will either pass through to the secured area, or be blocked and have to try again.

12 Jan. 09

Matthew Byrne

13

FYI Progress report

Appendix I - CODE
1 - Program Code
i = 1; j = 1; y = fopen('c0r0.end', 'rb'); x = fread(y, 'short'); [m,n] = size(x); fsamp = 8000; timeLength = m/8000; %About 1/2 a second framelength = 256; %overlapping nFrames = floor((m-128)/(framelength-128)); %fclose(x); figure(1); plot(x) sampNo = 1; for frame = 1 : nFrames, aquiredData = x(i:i+256-1); i = i + 128; figure(2); plot(aquiredData); %pause; % Hamming windowing windowing = (hamming(256).*aquiredData); figure(3); plot(windowing) % FFT FFTofData = fft(windowing); figure(4); plot(abs(FFTofData)) FFTofData = FFTofData(1:end/2); figure(5); plot(abs(FFTofData)) % First 128 frames of the sample W=melFilterMatrix2(8000,256,20); FilteredData = W*(abs(FFTofData)); figure (6); plot(FilteredData) MFCC = DCT(FilteredData); figure(7); plot (MFCC) end

12 Jan. 09

Matthew Byrne

14

FYI Progress report

- MFCC Code

function W = melFilterMatrix(fs, N, nofChannels) %for test, use these parameters %parameters fs = 8000; N = 256; nofChannels = 19; %compute resolution etc. df = fs/N; %frequency resolution Nmax = N/2; %Nyquist frequency index fmax = fs/2; %Nyquist frequency melmax = freq2mel(fmax); %maximum mel frequency %mel frequency increment generating 'nofChannels' filters melinc = melmax / (nofChannels + 1); %vector of center frequencies on mel scale melcenters = (1:nofChannels) .* melinc; %vector of center frequencies [Hz] fcenters = mel2freq(melcenters); %compute bandwidths %startfreq = [0 , fcenters(1:(nofChannels-1))]; %endfreq = [fcenters(2:nofChannels) , fmax]; %bandwidth = endfreq - startfreq ; %quantize into FFT indices indexcenter = round(fcenters ./df); %compute resulting frequencies %fftfreq = indexcenter.*df; %compute resulting error %diff = fcenters - fftfreq; %compute startfrequency, stopfrequency and bandwidth in indices indexstart = [1 , indexcenter(1:nofChannels-1)]; indexstop = [indexcenter(2:nofChannels),Nmax]; %idxbw = (indexstop - indexstart)+1; %FFTbandwidth = idxbw.*df; %compute matrix of triangle-shaped filter coefficients W = zeros(nofChannels,Nmax); for c = 1:nofChannels %left ramp increment = 1.0/(indexcenter(c) - indexstart(c)); for i = indexstart(c):indexcenter(c) W(c,i) = (i - indexstart(c))*increment; end %i %right ramp decrement = 1.0/(indexstop(c) - indexcenter(c)); for i = indexcenter(c):indexstop(c)

12 Jan. 09

Matthew Byrne

15

FYI Progress report

W(c,i) = 1.0 - ((i - indexcenter(c))*decrement); end %i end %c %normalize melfilter matrix for j = 1:nofChannels W(j,:) = W(j,:)/ sum(W(j,:)) ; end

12 Jan. 09

Matthew Byrne

16

FYI Progress report

Appendix II - TABLE OF REFERENCES

Frank J. Owens - Signal Processing of Speech http://www.lumenvox.com/resources/tips/HowLVsoftwareUsed.aspx http://www.blueface.ie http://www.speech-recognition.de/matlab-examples2.html http://tldp.org/HOWTO/VoIP-HOWTO-4.html http://www.ifp.uiuc.edu/~minhdo/teaching/speaker_recognition/ http://www.hindawi.com/GetArticle.aspx?doi=10.1155/S1110865704310024 http://cslu.cse.ogi.edu/HLTsurvey/ch1node9.html

12 Jan. 09

Matthew Byrne

17