User Manual
Contents
1. E 2 go Y E 2 g M3 gO E m3 O E 4 E O E O0 Bee fF amp amp Bee OF amp amp Figure 22 Marks selecting phonogram s areas with speaker s speech and noise 21 SPEAKER IDENTIFICATION 6 SPEAKER IDENTIFICATION The module identifies speakers with the help of the following methods 1 Pitch statistics method 2 Spectral and formant identification method SFIM 3 Total variability method TotV 4 The generalized method Speaker identification based on the first three methods involves two main steps speaker s voise modeling and decision making concerning identification The generalized method is the most accurate method it includes all three methods of speaker identification Models created with the help of the generalized method contain data with all three methods of voice identification in a single file The result of identification via the generalized method is an average identification result of methods mentioned above When using the generalized method FR and FA values are not calculated 22 SPEAKER IDENTIFICATION 6 1 Pitch statistics method This method uses sixteen different characteristics of voice pitch There are some of them the mean value of pitch its maximum and minimum values the median percentage of areas with increasing tone the variance of tone s logarithm the asymmetry of tone s logarithm the excess of tone s logarithm and other characteristics The v2. Al Plugins The set of modules for automatic identification STC S522 User s Guide Speech O Technology Center ABSTRACT Thank you for purchasing SIS Il Sound Editor We hope that our software will improve the quality of your tasks accomplishment Before getting started read this paper Al Plugins The set of modules for automatic identification STC S522 User s Guide then manual carefully This User s Guide is intended for operators who use the set of modules for automatic speaker identification by voice Al Plugins as a part of the specialized Sound editor SIS II It contains 1 General information about modules 2 Modules installation 3 Modules registration procedure 4 Estimation calculation of signal characteristics 5 The order of speech detection 6 The order of speaker identification 7 Trouble shooting There are the following data in the appendices A Explanations to qualitative characteristics of pbhonogram B List of terms and definitions C Abbreviations Any part of this publication may not be reproduced transmitted stored in a retrieval system or translated into any language in any form or by any means without the written permission of Soeech Technology Center Ltd CONTENTS AT oa A E E E E 5 PE E ueueass 4 Manpowertequite E aE nia ap ebro 4 TVPOgrapaYrCONVENTIONS veneran 5 CONAN adios 6 1 GENERAL INFORMATION sarria 1 1 About the DFOGUCE and the Produc usisni ai 7 12 Pro
3. APPENDICES A 4 Reverberation time Reverberation reverb is a complex blend of multiple interacting reflections within an enclosed space which combines with the direct sound from a source and defines the character of the sound in a room or hall Reverberation is the gradually reducing process of sound intensity at its multiple reflections Sometimes reverberation is called emulation of the effect with the help of reverberators Conditionally the accepted reverberation time is the time at which the reflected sound level is reduced by 60 dB Reverberation is type of the distortion where reflections of the source signal from any surfaces in a closed space and summing of these reflections with direct sound Reflected sounds come to microphone or to our ears with some time delay and some change in frequency components caused by distance from the signal source to the reflecting surface and this surface absorbtion characteristics In atmost manifistation it is echo repetition of a sound due to the reflection of sound waves unchanged spectrally To calculate the reverberation time we use the formula which invented by Sabin he was the first researcher of 0 164V A absorption A 4451 495 ai acoustic sound absorption coefficient it depends on the material its dispersed or frictional characteristics Si area of each surface architectural acoustics T where V the amount of space A genera l fund of acoustic
4. Click OK and make available the window with the appropriate type of the signal If speech is not segmented in the selected files for identification a warning message will appear Fig 29 as Waring speech is not segmented in A CifUsersstc Desktop signals Lava speech is not segmented in CifUsersstc Desktop signals 2 wav Continue identification Yes Figure 29 Warning about the presence of non segmented speech To improve the reliability of identification click No segment the speech using Speech detection method and repeat the identification process 30 APPENDICES APPENDICES Appendix A Explanations to qualitative characteristics of phonogram A 1 Irregularity of the frequency response Amplitude frequency characteristic AFC is a function that shows the frequency dependence of module of a complex function More often it means the modulus of a complex coefficient of linear four terminal device transmission It is also possible to consider AFC of other complex valued functions of frequency for example of the spectral density of signal power Amplitude frequency characteristic irregularity is the difference between maximum and minimum attenuation in the working bandwidth of shaping filter it is the degree of deviation from the straight parallel axes of frequencies Acoustics which have AFC as a straight line will be ideal for sound reproduction Unfortunately AFCs of real acoustic systems
5. Data Format Figure 7 Module options dialog box SIGNAL ANALYSIS CALCULATION OF SIGNAL CHARACTERISTICS 4 SIGNAL ANALYSIS CALCULATION OF SIGNAL CHARACTERISTICS Evaluation of signal characteristics is used to determine the suilability ofthe phonogram to the examination 4 1 Calculation of signal characteristics To evaluate calculate signal characteristics on the Modules menu click Signal Analysis Modules Signal Analysis The Signal Characteristics dialog box Fig 8 shows the main signal characteristics that determine the quality of the estimated phonogram amplitude frequency characteristic signal to noise ratio the presence or absence of tonal noise the average reverberation time and its dispersion the presence or absence of clippina the duration of pure speech ES signal characteristics 2 Spectral Characteristics 2 ARE Build graphic SMF 100 Build graphic Tonal Noise There are no stationary tonal noise detected Reverberation Time Average Time Dispersion Clipping There are no clipping detected Speech Total Llse channel Left Channel Right Channel Process Entire Signal 58 20 seconds Selected rea 58 66 seconds Visible part 68 20 seconds Save to project Copy Calculate Close Figure 8 Signal Characteristics dialog box before the calculating process 14 SIGNAL ANALYSIS CALCULATION OF SIGNAL CHARACTERISTICS To calculate si
6. sound 34 APPENDICES A 5 Clipping Clipping is an amplitude distortion For example it occurs when signal levels try to exceed the available amplitude range for given audio equipment The tops and bottoms of clipped waveforms are typically squared off generating frequencies that were not in the original signal Clipping is a form of waveform distortion that occurs in case of overloading of amplifier and exceeding the supply voltage limit by output voltage Subjectively it is expressed in the hissing and sizzle appearance Clipping occurs in digital processing when signal oversteps certain limited range For example in a 16 bit system the value of 32767 is the most positive which can be represented and if during the processing of the signal the amplitude will be doubled for example 32000 shoul be 64000 then instead the amplitude is truncated to the maximum 32767 As a result the top of the waves are flat and there are gross distortions of the signal 35 APPENDICES A 6 The duration of speech Speech is a concrete speaking occurring in time and clothed in sound including inner pronunciation or in writing Speech is the process of speaking and the result of this process i e speech activity voice work fixed by memory or a letter In this context speech and its duration are the areas segments of the phonogram that are meaningful in terms they contain only verbal information and its percen
7. SF 99 923 0 061 1233 49 DET sF 99 933 0 081 12353161 4 Pitch 99 599 4 050 1244991 DET z Pitch 99 599 4 050 1244991 DET DET iM Toti 99 920 0 0617 1241 052 DET Tot 99 44 0 06 1 141 351 Z Similarity 99 910 Speakers similarity ratio similarity 49 921 Speakers similarity ratio Compare Copy results Save to project Compare Copy results Save Eo project Save to project Close a speech is not segmented b speech is segmented Figure 24 Identification windows after performing the identification process Identification results can be copied to the clipboard by clicking the Copy button and pasted into a report or a text editor such as Notepad Fig 25 To obtain FR FA and DET graphs Fig 26 for each of the methods click the button PET in the field of the method If you select the Display EER check box the EER value will be shown in graphics The buttons make it possible to copy to the clipboard each of graphics in order to paste them into a text editor for a report 28 SPEAKER IDENTIFICATION Ej Untitled Notepad Le JE meza File Edit Format View Help comparison result File 1 C sUserssstcsDesktoprsignals 1 waw Right char Format 16 bit stereo 8000 Hz 88 20 sec speech segmented pure speech 13 51 sec File 2 c Users stc Dbesktop signals way Right char Format 16 bit stereo 8000 Hz 88 20 sec speech segmented pure spe
8. m 1 M SNR m where M total number of frames in the file 2 Integral SNR in frequency bands throughout the file SNR is calculated in frequency bands at time frames along all frequency bands ranging from zero frequency to Nyquist frequency SNR k m max 0 X k m N k m 13 Then integral SNR is calculated recursively in frequency bands throughout the file E SNR k 1 M gt m 1 M SNR k m where M total number of frames in the file Conversion of each type of SNR in decibles is carried out by a well known formula SNR GB 10 log SNR 32 APPENDICES A 3 Stationary tonal noise Noise is some extra signal added to useful signal for example any signal in the background of a speech conversation will be recorded as noise noise from the street hum and hindrances from home appliances electrical devices clicks in radio channel bangs of door closing rattle of paper another conversation music TV sound hindrances from the recording system or transmission channel itself etc In critical cases this strong additive noise may substitute the useful signal due to small dynamic range of the recorder and small sound level of speech Noise is an unwanted physical phenomenon or effect of electrical magnetic or electromagnetic fields electric currents or voltages of external or internal source that violates the normal operation of facilities or causes the degradation of technical characteristics and the par
9. tract of each person and this fact is reflected in different spectral characteristics of speech Difference of spectral characteristics is shown clearly in the frequencies orientation and mutual location of the formants In addition this method is based on the highlighting and comparison of location and behavior dynamics of three or more formants protected by Russian patent The usage of spectral formant method provides the value of EER up to 6 7 The value of this index for a particular case depends on the duration and quality of the speech fragments being compared Spectral formant method is the primary method These are the reasons the method makes lower requirements to the quality of the signal than other ones it is possible to operate with the signals that have the signal to noise ratio up to 10 dB the method shows relatively high rate of highlighting of soeech characteristics and it is enough resistant to the channel s type Speaker s voise modeling is as follows The average phonogram s spectrum which is used for the construction of identity speaker card is constructed Then for this phonogram the normalizing function is built it will be used to calculate instantaneous spectra of the speech signal Locations of three spectral peaks are used as the identifying features They are the most appropriate for formants of the speech signal at each spectral slice where it was possible to determine them reliably These charac
10. up or down from zero position Audio sound record phonogram Speech signal pre recorded in the file Data A graphical image in the info data window gathered while recording audio reading files operating with the program SIS II A representation of oscillograms waveforms spectrograms histograms and other graphical images Data box In SIS Il the independent rectangular area limited to a framework within the central working area of the main window of the program in which certain data oscillograms spectrograms histograms formants etc are displayed in the form of a graphic representation Data tab Independent data that together with other data is stored in one data window while operating with the program F Formant The amplitude maximum area of energy concentration in the speech sound spectrum determined by the resonant properties of the vocal tract In the speech sound 3 6 formants are commonly distinguished within the frequency range from 250 to 5000 Hz Formant is a phonetic characteristic of sound it contains information about the speaker s individual speech features Formant with the lowest frequency is denoted F the second F and so on to the highest frequencies 37 APPENDICES Fragment In SIS Il the part of data which is singled out in some way from the segment but has not lost its connection with the remaining data lt can be for example part of a segment limited by temporary marks or part o
11. If the necessary modules are not available in the list click the Refresh button If after this operation there are necessary modules missing again in the Plugins registration dialog box click the Paths button and in the Pathes to find modules dialog box Fig 5 click New Pathes to find modules pluging Me Delete OF Cancel Figure 5 Pathes to find modules dialog box In the Select folder dialog box Fig 6 specify the folder where the module is installed E select folder Look in G00 UE y amp My Computer R stc lt B E Contacts E Links BE Desktop A Music E Documents UE Pictures Je Downloads E sawed Games Videos e Signals R e i Favorites Je searches Directory Signals Files of type Directories Cancel Figure 6 Select folder dialog box In the Pathes to find modules dialog box select the necessary check box and click OK Maybe after this operation it will be necessary to click once again the Refresh button MODULES REGISTRATION If you select module s name and click Options in the Plugins registration dialog box you can become familiar with its properties in the Module options dialog box Fig 7 Module options Marne Signal Analysis Path C Program Files Speech Technology CenterS15 1 plugins Signal4nalys Sigral4n Identificator i 54E278F CO04 4075 8546 05CO006 2001 Description mee of signal characteristics
12. Typography conventions The following typographic conventions are used in the manual Font Description Normal Body text of the manual Italic The first appearance of a term Meaning of the term is explained here or in the appendix Also it is used to attract attention or to make up notes Bold Names of software components and interface elements headings buttons etc Boldltalic Names of files and paths to them Menu selection is marked with an arrow i e the combination Menu Command should be understood as following select Menu and then find the item Command To indicate the importance of any information the following comments and notes are used in the manual Jj Note Useful information x Caution Essential instructions which are obligatory to be fulfilled to prevent any fatal error in the system functioning INTRODUCTION Copyright SIS Il is trademark of Speech Technology Center Ltd All rights reserved All other companies and products mentioned in the manual are property of their respective owners The software includes modules of cross platform application framework Qt 4 7 0 htto at nokia com distributed under the terms of the GNU LGPL 2 1 license http www gnu org licenses Igpl 2 1 html GENERAL INFORMATION 1 GENERAL INFORMATION 1 1 About the product and the producer Name Conditional name Producer Postal address Telephone Fax The set o
13. VAD the areas of input signal are marked at which there is soeech and no noise clicks tone bursts etc Speech areas are cut to frames on which power spectrum of the signal is calculated Current estimates of power spectra of input signal X k m and noise N k m where k and m the indices of frequency and frame are calculated according to the power spectrum In this signal areas marked as non verbal are excluded from the analysis Then low informative and potentially dangerous in terms of possible noise bands are cut off below 100 Hz and above 3300 Hz and the ratio is calculated in the resulting band INR m 1 K x k 1 K X k m N k m J It represents the average value which shows to what extent the power spectrum of the input signal is more than the power spectrum of noise at a given time frame i e gives some general estimation of noise level at the frame The INR value m is compared with a threshold and only frames with INR m which are higher than the default threshold are sent for further processing The following characteristics are calculated 1 Integral SNR throughout the file In the frequency band 100 3300 Hz SNR is calculated at time frames SNR m lt X k m gt lt N k m gt 1 where lt gt averaging operation along all frequency bands Then SNR is stored recursively at time frames to obtain an estimate of the integral SNR through the file E SNR 1 M gt
14. alue of the equiprobable acceptance rejection error EER for the pitch statistics method depends on duration of the speech fragments being compared and can reach 18 19 The implementation of this algorithm is now realizable due to the creation by STC specialists the fully automatic high precision algorithm for pitch selection The advantages of this method are the following high rate of features comparison and as a consequence high speed of search or verification of speaker At the same time dependence of the reliability of this method on emotional and psychological state of speaker at the time of speech delivery is the cause to use it as an additional one Speaker s voise modeling is as follows Pitch extraction realizes via spectral analysis of phonogram s signal The method is based on algorithm that is based by turn on analysis of the values of pitch harmonics in the signal spectrum The algorithm is aimed at overcoming the problems associated with the suppression of half of signal spectrum in the channel For example the signal is missing in the telephone channel in the band from 0 to 300 Hz but the value of pitch frequency is in this very range In this case the algorithm of pitch allocation makes it possible to solve the problem by taking into account the harmonics of pitch frequency in the band from 0 3 to 3 4 kHz This type of identification is stable enough to signal to noise ratio If the level of the harmonics of pitch frequenc
15. ameters of these facilities Any oscillation in solids liquids ana gases can be the source of an audible and inaudible noise Radio electronic electromagnetic noise is a random variation of current or voltage in radio electronic devices for example audio recording and reproducing equipment Tonal noise TP is periodic signals with a frequency of the fundamental period from several tens of Hz and above Stationary noises are steady for whole recording or its fragment without any perceptible change in its characteristics Most background noises are more or less contituous street office hum hindrances from equipment without sharp changes in working modes etc Non stationary noises have breaks or pauses like beep signal in telephone channel Noise is primarily signals which spectrum overlaps with the PC spectrum and the amplitude of spectral components is comparable with the amplitudes of PC spectral components otherwise they are not noise Examples of tonal noise classes periodic pulse processes that generate a comb of harmonics of pitch in the spectrum tonal network noise 50 60 Hz and their harmonics PC acoustic noise peaks of spectrum in low frequency region tone pulses in telephone lines etc acoustic noise of artificial origin sirens music etc Comb filter A filter whose frequency response exhibits a series of deep peaks or notches equally spaced in frequency hence the word comb 33
16. are curves with many peaks and valleys Appearance of this irregularity when playing sounds of different frequencies is caused by no idealness of components as well as acoustical system in general the presence of various kinds of parasitic resonances vibrations of the shell etc The more uniform the AFC the more natural the reproduction The degree of irregularity of AFC is characterized by the ratio of peak value of sound pressure to minimum one or by other method the ratio of maximum minimum value to the average in a given range of frequencies expressed in decibles 31 APPENDICES A 2 Signal to noise ratio One of the key indicators affecting the quality of speaker identification the results of acoustic treatment etc is the original speech signal ratio to noise ratio the so called signal noise ratio SNR In this case a signal level is its capacity Signal to noise Ratio SNR is the ratio of the total signal to the total noise which shows how much higher the signal level is than the level of the noise A measure of signal quality Signal to noise ratio matrix is calculated in the Signal Analysis module SNR k m where k and m the indeces of frequency and frame on which the output parameters can be obtained They are the following 1 Integral SNR of the entire file in frequency bands 2 Integral SNR throughout the file SNR estimation algorithm is as follows With the help of Voice Activity Detector
17. ation For automatic speaker identification on the Modules menu click Speaker identification In the Identification dialog box Fig 23 from the drop down lists File 1 and File 2 select files to compare and select the check boxes methods of comparison in the Method field and click Compare To cancel the identification process click the Close button If the signal has not passed the segmentation procedure detection of speech Fig 23 a the warning will appear In the warning dialog box you should choose whether to continue identification or not If you choose Continue identification the identification module will produce the segmentation by itself ES Identification Fa Identification B mE Files File 1 An lav 16 bit mono 11025 Hz 200 57 sec speech not segmented File 2 A 4 mp3 16 bit mono 11025 Hz 200 57 sec speech not segmented Method SF 4 Pitch Copy results Save to project Close a speech is not segmented Files File 1 E law 16 bit mono 11025 Hz 200 57 sec speech segmented pure speech 95 09 sec File 2 A 4 mp3 Y 16 bit mono 11025 Hz 200 57 sec speech segmented pure speech 95 09 sec Method E 4 Pitch W Similarity Copy results Save to project Close b speech is segmented Figure 23 Identification windows before performing the identification process When conducting the speech segmentation procedure al
18. d noisy areas of phonogram on the Modules menu click Speech Detector Modules Speech detection in the Speech detection dialog box Fig 20 specify time interval between noisy areas where they will be merged into one ES speech detection Detect beeps Detect overloads Detect glitches Join noised intervals distance between them is less than 0 50 sec 0 1 sec E 1 sec Calculate Close Figure 20 Speech detection dialog box The program module carries out the search of speech segments and places them in the group of marks VAD In addition the program module can carry out the detecting of beeps overloads and glitches Marks of the given kinds of signals are located in the general group of marks Noise For detecting of beeps overloads and glitches select the necessary check boxes Detect beeps Detect overloads Detect glitches To start the speech detection process click the Calculate button To cancel the speech detection process click the Close button The calculation process speech detection takes time it is displayed in the Task Viewer dialog box The process can be interrupted by clicking the button to the right of the operation s progress indicator SPEECH DETECTI N 5 2 Speech detection s results After speech detection the interval marks will appear in the data window they mark phonogram s areas with speaker s speech and noise The intervals will be shown in
19. duc GUO CAM OI misa ii E O 8 ES Compositon ana tacles asasarsnaca reia 8 2 INSTALLATION TACSET OF MOLULE eniinn a naa E AN E EAE AOA 9 MODULE SREGDITRA ON sanan enanas 11 4 SIGNAL ANALYSIS CALCULATION OF SIGNAL CHARACTERISTIC S ssssssssssssssssssseserersssssssseeerrrrsssssssssrerrreses 14 4 1 Calculation of signal CharacteristiCs cssssssssssscssssssssscsessssescscsescsesesescsesesesesesssesssssssssssssessssssessseseseseseseseseeseseseseees 14 4 2 Graphics of spectralcharacten UGS ssi cnmaiaamasaicn rato 16 4 3 Estimations and recommendations on the received signal characteristicS ocoocococononononororororornrnonononoos 17 4 4 Copying of signal characteristics into a text editor ssesssessesseessessesssessesseessesseessesseoseessesseesersseosersseoseossesseoseesses 18 O A 19 SEVEN KE CU CL ON is a 19 9 2 SPEECH detection Ss tesUliS siones EEEE 20 ss O A e e EII 22 6 1 Pen SEMIS EI GS TMU VO siers eesi rnn ias 23 6 2 Spectral and formant identification method seessessesseseesessessesseseesessessesseseeseosessesseneeseosesseneeseeseosesseneeseeseoseseseeseess 24 Sek E E TANS UI Os italia A E 25 6 4 Theoretical basis of the generalized method ooocococococococococonononononononcncncnonononono nono no nono nono no nn i iai iiaa 26 gt SU OF Ns the Jaentl cats sonal A 27 A A E e O OR 30 FAV a EEAS Ee E A E E A N A A E E A 30 a aa Ma E ENET AANE NEN EESE ENA A 3 Appendix A Explanations to qualitative character
20. e 10 The spectrum of the Fast Fourier Transform FFT is build for AFC This graphic is shown in Figure 11 AFC boundaries are defined by module with the dashed lines i olg Ee Manager Panel Windows Marks Projects DB B 9 1 FFT Power Figure 11 FFT spectrum for AFC Frequency distribution is build for the signal to noise ratio Fig 12 oe 1000 1500 000 S00 S000 3500 4000 4500 S000 Figure 12 Window with SNR and the selected new fragment The boundaries within which the integral value of SNR is calculated are marked with the dashed lines as given in the Signal Characteristics dialog box By default the integral value of SNR is calculated in the range from 100 to 3300 Hz If selecting another fragment on the SNR graphic Fig 12 frequency value is changed and the SNR integral value is recalculated in the Signal Characteristics dialog box Fig 13 SMR 1191 1906 Hz 29 dE Figure 13 Selected frequency fragment and integral value of SNR for this fragment SIGNAL ANALYSIS CALCULATION OF SIGNAL CHARACTERISTICS 4 3 Estimations and recommendations on the received signal characteristics To obtain recommendations on each group of characteristics click the button J Examples of recommendations to signal characteristics in Figure 10 are shown
21. ech 13 51 sec FR FA LR 994 948 0 081 1233 357 959 575 0 080 1244 685 566 893 0 081 1240 674 similarity 99 B83 I Figure 25 Result of comparison copied into a text editor Notepad ES FR FA DET Plot P fete FR FA DET FR FA plot for comparison DET plot for ee ee ee ee ee Cee mm mmm U M AAA AE gt A ERE E A A A O A a Oo u a a u n pe Pp pp a Gna AAA AA pp CEE FR FA 90 a o la a y A a l u A oa A ee ee ces mes ee do A DA AO TA CA A A oe m e aL 1 05 0 05 1 15 2 25 O 20 40 60 0 100 Distance FR 00 Copy FR FA Display EER Copy DET Figure 26 Examples of FR FA and DET graphs 29 TROUBLE SHOOTING 7 TROUBLE SHOOTING 7 1 Warnings and Errors If while selecting the calculating of signal characteristics an active window has not got an oscillogram of this signal an error message will appear as shown on Figure 27 E Signal characteristics x Active signal isn tany sound type Figure 27 Error message of data choice to calculate characteristics Click OK and make available the window with the appropriate type of the signal If in order to detect speech a window that does not contain an oscillogram of the signal is chosen a warning message will appear Figure 28 speech detection i Active signal iint any sound type Figure 28 Warning about the absence of the required type of the signal
22. ened contextual menu recommendations to calculate the duration of speech Data appearing in the Recommendations windows can be copied to the clipboard to paste subsequently into the report To copy the data click the right mouse button to invoke the contextual menu Fig 18 or press Ctrl C 17 SIGNAL ANALYSIS CALCULATION OF SIGNAL CHARACTERISTICS 4 4 Copying of signal characteristics into a text editor To put the calculated signal characteristics to the clipboard of a text editor click the Copy button in the Signal Characteristics dialog box In the text editor perform the Paste operation Example of signal characteristics placed in the text editor is shown in Fgure 19 In the Speech item the duration of pure speech without noise and pauses is depicted lal Untitled Notepad File Edit Format View Help speech For a Ea value of speech duration EER 15 Total solution 10 51 Cepstrum method 11 18 Spectrum Tormant method 33 57 Pitch method 35 13 4 il Figure 19 Signal characteristics in the text editor Notepad SPEECH DETECTION 5 SPEECH DETECTION Detecting of speech is required for preliminary segmentation of phonograms coming for processing in the speaker identification module Thus the phonogram is divided into areas segments with a useful speech signal noise pauses and telephone signals 5 1 Detection execution To extract speech automatically from the backgroun
23. f a segment included in the highlighted interval between permanent marks or part of a segment visible in the box M Mark A tool to highlight specific data areas in the data window N Noise 1 Disorderly oscillations of a different physical nature having continuous spectrum in a sound frequency range 2 Unwanted sound that complicates the useful signal determination and use Any oscillation in solids liquids ana gases can be the source of an audible and inaudible noise Radio electronic electromagnetic noise is a random variation of current or voltage in radio electronic devices for example audio recording and reproducing equipment Normal distribution mixture A general linear combination of Gaussian functions used for approximation of various experimental distributions of the acoustic soace components O Operator A person who uses the program as intended Pause lat pausa gr pausis stop termination A break in speech which acoustically corresponds to the absence of sound and physiologically to the stop in the activity of soeech organs Pitch fundamental frequency pitch of sound voice A perceived quality of sound that is most closely related to the frequency of the first harmonic fundamental frequency in a discrete spectrum and depends on the size and speed of vocal cords vibrations In oral speech this feature determines voice type bass tenor descant etc Pitch of voice sound A property of vo
24. f modules for automatic identification Al Plugins S1E55Z2 Speech Technology Center Ltd Russia 196084 St Petersburg 4 Krasutskogo str 7 812 325 88 48 7 812 327 9297 GENERAL INFORMATION 1 2 Product allocation The set of modules for automatic identification Al Plugins then Al Plugins or the set of modules as a part of the specialized Sound editor SIS Il is intended for an estimation of characteristics of a signal detecting of speech and noise and speaker identification by voice 1 3 Composition and facilities The set includes the following additional program modules 1 Signal Analysis The quality estimation module of a phonogram allows estimating automatically the sullability ofthe phonogram to the examination 2 Speech Detector The search module of speech segments allows allocating automatically speech and noise segments in a phonogram 3 Speaker identification The identification module carries out automatic speaker identification by voice using three methods and the generalized solution lt should be noted that features of the program are constantly increasing and improving so it s recommended to specify the current additional modules on the STC official website http www speechpro com or please contact Speech Technology Center managers to find out more about INSTALLATION OF THE SET OF MODULES 2 INSTALLATION OF THE SET OF MODULES Db The set of modules should be installed on a PC whic
25. gnal characteristics specify the process area Entire Signal Selected Area Visible part and click the button Calculate The calculation process of signal characteristics takes time it is displayed in the Task Viewer Fig 9 dialog box The process can be interrupted by clicking the button to the right of the operation s progress indicator Task Viewer Signal characteristics 1 x x Figure 9 Indication of the operation Calculation s results are represented in the Signal Characteristics dialog box Figure 10 ES signal characteristics 2 Spectral Characteristics A AFC 04 3672 H Build graphic SWR 100 3300 H 37 dB Build graphic Tonal Noise gt There are stationary tonal noise detected Total 5 57 Sec Reverberation Time Average Time 312 mec Dispersion 24 Clipping gt There are clipping detected Total 9 00 Sec Total 10 20 o Speech sa Total 15 75 Sec Total 17 56 Use channel Left Channel d Right Channel Process Entire Signal 58 20 seconds Selected rea 5 66 seconds Visible part 88 20 seconds Save bo project Copy Calculate Figure 10 Signal Characteristics dialog box after performing the calculations 4 2 Graphics of spectral characteristics To build the graphics of the amplitude frequency characteristic AFC of the signal or the graphics of the signal to noise ratio SNR click the button Build graphic in the AFC or SNR fields refer to figur
26. h already has the specialized Sound editor SIS II No additional maintenance or software to install or modules to operate is required The set of modules should be installed as an addition to the specialized Sound editor SIS II To start the installation run the file AI_Plugins exe Further steps to install the set of modules are shown below In welcome window Fig 1 click Next gt and follow the instructions of the Installation Wizard appearing on the screen fa STC SIS I Automatic Identification Plugins 112 739 Setup bo peel C J oad Welcome to the STC SIS II Automatic Identification A Cantar gy Plugins 1 12 739 Setup Wizard The Setup Wizard will install STC STS 11 Automatic Identification Plugins 1 12 739 on your computer Click Mest to continue or Cancel to exit the Setup Wizard WARNING This program is protected by copyright law and international treaties Figure 1 Welcome window INSTALLATION OF THE SET OF MODULES On completion of the installation click the Finish button to exit the Setup Wizard program Fig 2 Se na G oo e an ES A E omg STE SIST Automatic Identification Plugins 112 739 Setup col El aa Technology Plugins 1 12 739 Setup Wizard speech Completed the STC 515 I Automatic Identification Center Click the Finish button to exit the Setup Wizard Ce CE Figure 2 Window of successful setup installation MODULES REGISTRATION 3 MODULES REGISTRATION Reg
27. h is lower in frequency than 17 Hz is called infrasound while ultrasound is an oscillation with a frequency greater than the upper limit of human hearing 20 000 Hz Speaker A person whose speech is in an audio sound record Speech sound A minimum unit of speech flow resulting from human articulation activity Speech sound is characterized by specific acoustic and perceptive properties V Voice Activity Detection VAD Software tool to separate active speech from background noise or silence W Waveform oscillogram Waveform of the speech signal is a graphic representation of the signal vibration amplitude as a function of time Waveforms can be obtained using signal processing equipment loop waveform viewers signal level recorders and electronic waveform viewers Waveforms can be used to extract fragments of data for further research 39 APPENDICES Appendix C Abbreviations In the present paper the following abbreviations are used DET Detection Error Trade off Plo graph of the relation of errors of classification which visually shows the dependence of value of probabilities of type errors and type ll errors DRR Direct to Reverberant Ratio intensity ratio of direct sound to the reverb sound ratio EER Equal Error Rate level of an equal error or point of coincidence of probabilities of type I errors a errors false positives and type Il errors B errors false negatives FA False Acceptance false rec
28. ice measured by the vocal folds oscillation frequency in a unit of time the more oscillations account for a unit of time the higher is the pitch Range A quantity setting the utmost limits of attribute change e g sounding speech attributes difference between minimum and maximum values of the attribute 38 APPENDICES Speaker identification by voice Algorithm of calculation allocation of speech identifiers ID and algorithm of paired comparison of them The process of comparing the speech of an unknown speaker against a database of the speech samples of known speakers to determine whether it matches any of the templates or not i e to identify the submitted unknown speaker with any of known speakers Speaker identification characteristics The stable individual characteristics of a speaker that are obtained from his speech appearance and speech characteristics as well as subjective auditory estimation of a speaker Sound spectrum An acoustic representation of complex sound providing information about the frequency of sound source pitch harmonics and relative intensity of all its frequency components Sound A mechanical oscillation travelling through elastic mediums or bodies solids liquids and gases composed of frequencies within the limits of human hearing between about 17 20 Hz and 20 000 Hz The heightened sensibility of human ear is detected in the frequency range from 1 kHz to 5 kHz Mechanical oscillation whic
29. in Figures 14 18 os E5 Recomendations Spectral Characteristics For a specified SMR value EER is Total solution 2 40 Cepstrum method 3 24 Spectrunm Formant method 12 17 Pitch method 20 95 Figure 14 Recommendations to calculate spectral characteristics 2 Recomendations Tonal Noise If tonal noise is detected in a signal you are recommended to perform noise reduction r a A ES Recomendations Reverberation Time For a specified value of reverberation time EER is Total solution 1 84 Cepstrum method 3 64 Spectrunm Formank method 11 17 Pitch method 21 26 Figure 16 Recommendations to calculate reverberation ES Recomendations Clipping If there are clipped segments in a signal you need to apply the Speech Detector plug in If max distance between clipped segments in a signal is less than 3 sec then automatic speaker identification cannot be performed Figure 17 Recommendations to calculate clipping ES Recomendatior5 Speech For a specified value of speech duration EER is Total solution 10 51 Cepistrum method 11 10 Spectrum Formant method 33 57 Pitch method 35 13 If max distance between clipped segments ina signal is less than 3 sec then automatic sp Z Ctrl C identification cannot be performed PY ii Select All Ctrl 4 Close Figure 18 Recommendations window with the op
30. istics Of phonogramM esessessessessessesseoseeseeseeseeseeseeseeseessessesses 31 Appendix B The list of terms and CePFINITIONS ce eesssssssssssesesescsesssssssssssssssesesecesesesesesesssseseseseccscseseseseseseseseeseseeeacens 37 POO MVCN Ce FE VIO Sn ruinas 40 INTRODUCTION INTRODUCTION General The given User s Guide discovers necessary data on installation adjustment and operation with the set of modules for automatic speaker identification by voice Al Plugins as a part of the specialized Sound editor SIS II This paper SIS Il the specialized Sound editor STC S521 User s Guide is intended for operators office workers who use the specialized Sound editor SIS Il according to its intended purpose The manual contains information how to install and operate the Sound editor SIS Il Thus it characterizes the possibilities ofthe solution and also describes the sound editor algorithm This paper does not replace academic reference books and manuals from the manufacturers of the operating system and common software Manpower requirements Staff producing the installation of the specialized Sound editor SIS Il should have professional skills to install general and special software Staff working with the dedicated Sound editor SIS Il should have basic skills to operate with applications in the operating systems Microsoft Windows and should know how to expertise speech audio sound records INTRODUCTION
31. istration of additional modules in the specialized Sound editor SIS Il is performed automatically After installing the set of modules and running the Sound editor SIS II these additional modules will be added to the Modules menu Fig 3 Register Formant Extractor speaker identification speech Detector Pitch Extractor signal Analysis Figure 3 Modules menu If the installation of modules was performed by one user and a new user operates with them the modules for the new user will not be registered automatically so it must be done manually On the Modules menu click Register In the Plugins registration dialog box Fig 4 all the installed modules will be depicted please select the check boxes Plugins registration Formant Extractor Pitch Extractor Signal Analysis Speaker identification Speech Detector Paths a ce ee Refresh OK Cancel Figure 4 Plugins registration window Additional software modules are executed as plug ins independently complied program modules that are dynamically connected to the main program By default they are installed in the Plugins located in the directory C Program Files Speech Technology Centen SIS INPlugins There are the following modules 1 Signal Analysis module SignalAnalys dll 2 Speech Detector module MarkingPlugin dll MODULES REGISTRATION 3 Speaker identification module IdentificationPlugin dll
32. l noise will be extracted as if all the check boxes were selected in the speech detection module refer to figure 20 The result of noise detection outside the module is not given and it can not be checked or updated by an operator As a result there will not be enough pure speech for identification In this case please cancel the identification and at first run the speech segmentation procedure using Speech detection module 27 SPEAKER IDENTIFICATION The identification modeling process takes time it is displayed in the Task Viewer dialog box and at the header ofthe Identification dialog box The process can be interrupted by clicking the button to the right of the operation s progress indicator Quality of compared phonograms is estimated by the module during the identification process After completion of identification results for the selected methods will be displaed in the table of the Identification dialog box Fig 24 3 Identification 2 ES Identification Files Files File 1 B 4 mp3 File 1 A 4 mp3 16 bit mono 11025 Hz 200 57 sec 16 bit mono 11025 Hz 200 57 sec speech nok segmented pure speech 46 50 sec speech segmented pure speech 95 09 sec File 2 14 5 mp3 File 2 B 5 mp3 16 bit mono 11025 Hz 200 57 sec 16 bit mono 11025 Hz 200 57 sec speech nok segmented pure speech 46 50 sec speech segmented pure speech 95 09 sec Method FR FA LA DIET Method FR FA LA
33. ognition FR False Rejection false access control LR Likelihood Ratio evaluation of relations of maximum likelihood functions maximum likelihood evaluation JFA Joint Factor Analysis MFCC Mel Fourier Cepstrum coefficients SVM Support Vector Machine VAD Voice Activity Detection and Silence Suppression voice activity detector is used for extracting of active speech from background noise or silence AFC Amplitude frequency characteristic SNR Signal to noise ratio NDM Normal distribution mixture SFIM Spectral and formant identification method 40
34. ort Vector Machine well established in terms of speed and quality of identification solution is used Type errors and type ll errors FR FA are calculated on the resulting SVM distance Bagging is performed on the duration of speech signal FR FA evaluations depend on the duration of speech signal on which bases every compared NDM module is built 25 SPEAKER IDENTIFICATION 6 4 Theoretical basis of the generalized method The generalized method of speaker identification is realized on the basis of the generic solution that is made according to the identification results by one or more methods of identification regardless of their number The result is a pseudo likelinood P belonging of two compared phonograms to one speaker which lies in the range from O to 100 0 means the minimum possible similarity of speakers in phonograms 100 the full compliance of speakers voices The algorithm of calculating of the generalized solution is based on the usage of the weighted voting method N P w f F oR Dd i l where N the number of authentication methods according to which the decision is received W weighting factor of L method The value of weighting factor changes automatically depending on signal quality FRIFA type errors and type ll errors of 1 method The FR F4 values are calculated as a persantage and have a range of values from O to 100 26 SPEAKER IDENTIFICATION 6 5 Performing the identific
35. tage with less significant nonverbal information eg noise This characteristic of the speech signal is fundamental in terms of the importance of speech identification as it directly related to the quantity of realizable acoustic and linguistic features in the useful signal 36 APPENDICES Appendix B The list of terms and definitions A Acoustic and phonetic attributes of oral speech The attributes reflecting acoustic qualities of the vocal tract and articulation skills of the person These attributes are perceived and revealed with the help of technical means and form the basis of instrumental analysis of speech signals the attributes can be evaluated quantitatively Active tab Tab of active data window used as a data source The tab is usually displayed over other tabs Amplitude magnitude lat amplitudo size The maximum deviation value from the equilibrium position of an oscillating quantity for example the deviation from zero of an in circuit electric current voltage sound pressure intensity etc It represents the size of vibration deviation value In strictly periodic vibrations the amplitude is a constant In the research of harmonic sound vibrations the amplitude means sound pressure in a signal expressed by the amplitude of a current voltage or other electrical quantity on the output of sound converting equipment microphone In the signal waveform figure the amplitude represents the deviation size of an image
36. teristics determine mostly individual features of vocal tract Density of the distribution of identifying features is modeled with using NDM Immediate decision making concerning identification is performed with using SVM classifier Support Vector Machine Type errors and type ll errors FR FA are calculated on the resulting SVM distance Bagging is performed on the duration of speech signal FR FA evaluations depend on the duration of speech signal on which bases every compared NDM module is built 24 SPEAKER IDENTIFICATION 6 3 Total variability method Total variability method TotV method is the most advanced method of speaker identification by voice Speaker s voise modeling is as follows Speech characteristics MFCC Mel Fourier Cepstrum coefficients are calculated They are features that describe spectral representation of the speech signal at different periods of time Density of the distribution of identifying features is modeled using Normal distribution mixture NDM module parameters with the help of specially adapted factor analysis are presented in the form of i vector of small dimension in the so called total variability space that includes subspaces of its own channels and its own voices used in JFA Joint Factor Analysis A distinctive feature of this representation of NDM module is its high self descriptiveness and the small size of data At the stage of comparing of i vectors SVM classifier Supp
37. the list of marks in the VAD and Noise tabs Fig 21 accordingly An operator can manually adjust the intervals by means of the Sound editor SIS II In this case speech areas given by the operator will be used for identification A Sao mI F Single Sounds Noise Speakers YAD Group Text Begin End Duration E yan 00 24 56 00 38 21 00 1364 V vac 00 13 72 00 14 83 00 01 11 M Search in marks O from O dk E Single Sounds Noise Speakers VAL o E Group Text Begin End Duration W J Noise 00 24 56 00 38 21 00 13 64 Y B noise 0013 72 00 14 83 00 04 11 4 Search in marks OfromO amp a a Figure 21 Data window VAD and Noise tabs after speech detection 20 SPEECH ODETE TON Overall number of marks and their length can be obtained by selecting the VAD and Noise groups on the Marks tab of the Manager Panel Fig 22 Manager Panel x Manager Panel x Marne Cole keys sel Wis Wis Cole Keys Sel Wis Vis T Single Ins Ho E Ins Ho E WB sounds E lt keyo gt E E skew gt HO E Moises E e OD Ts E lt ko O w amp El g E O E i Oo RH E i gE
38. y is at least a few decibles higher than noise level the algorithm by means of special underlining methods of pitch harmonics makes it possible to determine the desired frequency At the calculating stage of pitch statistical characteristics a set of values of pitch statistics according to the obtained pitch curves is determined This set includes the following characteristics the average maximum and minimum pitch values frequency values above which there is 3 of pitch values and frequency values below which there is 1 of pitch values the median percentage of areas with the increasing pitch the dispersion the asymmetry of logarithm and the excess of pitch logarithm the mean rate of pitch change and others At the decision making stage of identification pitch statistical characteristics are compared values of type errors and type ll errors FA and FR for each of the obtained values of pitch statistics are calculated as well as FA and FR values on the resulting metrics of pitch statistics as a weighed sum of all relative deviations of values of pitch Statistics for two compared phonograms are calculated Bagging is performed on the duration of speech signal FR FA evaluations depend on the duration of speech signal on which bases every compared NDM module is built 23 SPEAKER IDENTIFICATION 6 2 Spectral and formant identification method This method is based on the thesis about uniqueness of the geometry of the vocal
Download Pdf Manuals
Related Search