VocalTractLab 2.1 User Manual
Contents
1. Coarticulatory effects of the vocal tract are modeled using a coarticulation model based on bilinear interpolation as described in Birkholz 2013 In short for each consonant three vocal tract target shapes were defined in VTL they were extracted from dynamic MRI data one in the context of each corner vowel a i u For example for the consonant d we can denote these three shapes by d a d i and d u When d has to be realized in the context of an arbitrary vowel V in the simulation the coarticulation model maps V into the parameter subspace of the vocal tract shapes for a i u The position in the subspace is then used to infer the coarticulated target shape of d in the context of V by bilinear interpolation between d a d i and d u For example when the vowel e is half way between a and i then the target shape for d in the utterance ede is assumed to be half way between d a and d i Connected utterances in VTL are defined by gestural scores a concept taken from articulatory phonology Browman and Goldstein 1992p A gestural score consists of a number of independent tiers populated with discrete gestures Each gesture represents the movement of certain articulators i e vocal tract parameters toward a target configuration The change of vocal tract parameters in response to these dis crete gestures is governed by linear dynamical systems Details of the definition of gestural scores2. 100 200 00 1 eo e Position 0 106 s Area 0 004 Reset V Main path ese IV Side paths 200 00 4 Continue synthesis Time signal display Synthesis type 400 00 5 C Subglottal impedance E p y T y F C Supraglottal impedance 0 05 0 04 003 0 02 0 01 C Transfer function j E 4 Vowel LF flow pulse Phone full simulation 200 5 Gestural model F Choice of synthesi a a Set area function is type i LF glottal flow pulse E 4004 Acoustics time domain Bl 14 412 10 8 6 4 2 0 2 4 6 8 SS EI cm 2 J Selected section 24 P L 18 cm 14 12 18 16 El 24 22 20 Figure 13 Time domain simulation page tract model The spectra resulting from the synthesis with this and the previous two options can be shown in the spectrum display of the vocal tract page when the checkbox TDS spectrum is checked e Vowel LF flow pulse is used for the synthesis of a voiced sound excited by the LF flow pulse model Click the button Set area function to set the area function used for the synthesis to the one of the vocal tract model e Phone full simulation is used for the synthesis of a phone with a static vocal tract shape excited by a model of the vocal folds Set the area function with the button Set area function and set a rest shape of the vocal fold model used for the synthesis see below Depending on the shape of the glottis
3. Figure 16 Segment annotation dialog model FO curve as a red dashed line in the spectrogram of the main track to compare it with the measured FO contour of the master signal in the extra track 6 4 Export possibilities From a gestural score you can not only generate the speech signal but also save a sequence of video frames of the vocal tract and the trajectories of the virtual EMA sensors attached to the vocal tract model Saving video frames might be interesting if you want to create a video of the speaking vocal tract To save video frames click Export Vocal tract video frames from ges score in the menu You are then asked to select a folder Into this folder video frames of the vocal tract will be saved as BMP files with a frame rate of 30 Hz The look of the vocal tract in the images corresponds to the settings in the vocal tract dialog for example a 2D view or a 3D wire frame representation of the vocal tract These frames can then be compiled into a video file using for example the software VirtualDub www vixtualdub org With the menu item Export EMA trajectories from gestural score you can save the trajectories of the virtual EMA sensors to a plain text file The x and y coordinates of the individual sensors will be represented in different columns at a frame rate of 200 Hz 7 Some typical uses 7 1 Analysis of the glottal flow for different supraglottal loads 1 Select the time domain simu
4. PPT ORA dut ES Extra oscillo Hi iint eanan E bal Extra track oscillogram and spectrogram HOP OIE TNs MOTO stokes ich i Badin Prime mark JAA VAN UU Ti k I Show animation wa ma a dl M Badan Synthesize 4 Audio signals Extra spectro M Show extra track I Show sonagram IV Show segmentation IV Show model FO curve SEES Vowel gestures Show vocal tract Show glottis Lip gestures Calculate FO Tongue tip g tpostalveol Calc voice quality Analysis results Tongue body g Analysis settings Annotation dialog Velic gestures Selected gesture FO gestures Lung pressure g 1000 Pa 1000 Pa Figure 15 Gestural score page The gestural score page allows you to create gestural scores for the synthesis of connected utterances 17 The page has a control panel at the left side and a signal display and gestural score editor at the right Fig 15 The signal display allows the side by side comparison of the synthetic speech signal and a natural speech signal and so helps to reproduce natural utterances by articulatory speech synthesis 6 1 The concept of gestural scores A gestural score is an organized pattern of articulatory gestures for the realization of an utterance This concept was originally developed in the framework of articulatory phonology Browman and Goldstein 1992p While the basic idea is the same in
5. VTL the specification and execution of gestures differ from articulatory phonology and will be briefly discussed here In general a gesture represents unidirectional movement toward a target configuration of the vocal tract model or the vocal fold model by the partici pating articulators parameters These gestures are organized in eight tiers as shown in Fig 15 Each tier contains gestures of a certain type From top to bottom these are vowel gestures lip gestures tongue tip gestures tongue body gestures velic gestures glottal shape gestures FO gestures and lung pressure gestures Within the same tier the gestures grey and white boxes form a sequence of target directed movements towards consecutive targets Some tiers have the exclusive control over a set of vocal tract or vocal fold model parameters while other parameters are affected by gestures on different tiers Glottal shape gestures FO gestures and lung pressure gestures control the parameters of the selected vocal fold model The lung pressure gestures and the FO gestures exclusively control the corresponding parameters of the vocal fold model Here each gesture specifies a target value for the controlled param eter These target values are sequentially approached For FO gestures the target does not need to be constant in the time interval of a gesture but may vary as a linear function of time i e it may have a slope This corresponds to the target approximation model for FO c
6. Yi Xu for testing the software proofreading this manuscript and general feedback Parts of the research leading to the development of VocalTractLab were funded by the German Research Foundation DFG grant JA 1476 1 1 References Birkholz Peter 2005 3D Artikulatorische Sprachsynthese Logos Verlag Berlin 2007 Control of an Articulatory Speech Synthesizer based on Dynamic Approximation of Spatial Articulatory Targets In Interspeech 2007 Eurospeech Antwerp Belgium pp 2865 2868 2013 Modeling Consonant Vowel Coarticulation for Articulatory Speech Synthesis In PLoS ONE 8 4 e60603 DOI 10 1371 journal pone 0060603 URL http dx doi org 10 1371 2F journal pone 0060603 submitted Enhanced area functions for noise source modeling in the vocal tract In submitted Birkholz Peter and Dietmar Jack l 2004 Influence of Temporal Discretization Schemes on Formant Frequencies and Bandwidths in Time Domain Simulations of the Vocal Tract System In Inter speech 2004 Jeju Island Korea pp 1125 1128 Birkholz Peter Dietmar Jack l and Bernd J Kr ger 2006 Construction and Control of a Three Dimensional Vocal Tract Model In International Conference on Acoustics Speech and Signal Processing ICASSP 06 Toulouse France pp 873 876 Birkholz Peter and Bernd J Kr ger 2006 Vocal Tract Model Adaptation Using Magnetic Resonance Imaging In 7th International
7. constant in a fast approach To find the right time constant for a gesture is somewhat tricky but a value of 15 ms for supraglottal gestures has proven to give satisfactory results in most cases The value of a gesture is either a numeric value the subglottal pressure in Pa for pressure gestures the FO in semitones for FO gestures or the velum position for velic gestures or a label For a glottal shape gesture the label specifies a pre defined glottal shape as the target for the gesture For a vowel lip tongue tip or tongue body gesture the label specifies a pre defined vocal tract shape as the target for the gesture FO gestures can have a non zero slope in addition to the target value The duration of a gesture can also be changed by dragging the vertical border line between two gestures with the LMB When Shift t is pressed at the same time only the border is moved Otherwise all gestures right from the border are moved along with the border For velic gestures FO gestures and pressure gestures the numeric value of the gesture can also be changed by dragging the horizontal dotted line in the gesture box vertically with the LMB With a right click in the gestural score the red time mark is set to the position of the mouse cursor the gesture under the mouse cursor is selected and a context menu is opened From the context menu you can choose to insert a gesture or to delete the selected gesture Press LMB to set the time mark t
8. of your choice as numbers in a text file This allows the samples of a signal to be easily imported into programs like Matlab or MS Excel If the sampling rate of a WAV file to be loaded does not correspond to 22050 Hz the sampling rate is converted automatically There are four displays in the main part of the signal page cf Fig 1 the detailed oscillogram the overview oscillogram the spectrogram and the short time Fourier transform With the checkboxes next to the oscillogram display you can select the track s to display For each track the time signal is displayed in a different color The detailed oscillogram shows a detailed view of the highlighted part in the middle of the overview oscillogram For the track selected next to the spectrogram display the spectrogram is shown Both the overview oscillogram and the spectrogram have the same time scale You can change the scale with the buttons and below the checkboxes next to the detailed oscillogram With the buttons and that are arranged vertically you can adjust the amplitude scaling of the signals for display With the buttons and next to the word Main you can actually increase or decrease the signal amplitude in the main track However when a signal has small amplitudes it is usually better to select the menu item Edit Normalize amplitude or press the shortcut to increase the amplitude The buttons Main lt gt E
9. other hand the control parameters define the instantaneous shape and properties of the vocal folds e g the vocal fold tension in terms of the fundamental frequency and the degree of abduction In the dialog the control parameters are changed with scrollbars Individual settings of these parameters can be saved as shapes For example you can define a shape for modal phonation a shape for breathy phonation more abducted vocal folds in the pre phonatory state and a shape for voiceless excitation strong abduction These shapes are referred to in gestural scores where they are associated with glottal gestures Sec 6 The current list of shapes for a model can be found in the drop down list with the label Shape The buttons around the drop down list allow saving the current setting of control parameters as one of the existing shapes or as a new shape and to remove shapes from the list Please note that the control parameters FO and Subglottal pressure which exist for all vocal fold models are not included in the definition of a shape because these parameters are controlled by independent tiers of the gestural score The parameters of the vocal fold models and the associated shapes are stored in the 16 speaker file JD2 speaker which is loaded automatically when VTL is started Therefore to save any changes made to parameters or shapes you must save the speaker file by pressing or selecting File Save speaker
10. synthesis Help Pbi x Signals Vocal tract Time domain simulation Gestural score Show vocal tract a Length 16 14 cm A 1 55 cm 2 C 2 2 cm Velum area 0 000 cm 2 Velum pos 7 31 cm Tongue Vocal tract shapes 124 Cur area 1 557 cm 2 Teeth pos 15 44 cm Phonetic parameters 114 Min area 0 340 cm 2 ati2 0 cm LF glottal flow pulse 104 E 34 Area function display Cross section through the vocal tract Improve formants Check vowel shape Get formant error Edit pole zero plan Map shape to vowel subspace CV transition dialog Play short vowel 0 0 2 0 4 0 6 0 8 0 10 0 12 0 14 0 16 0 cm pl I Discrete V Area Circ IV Show branches V Text Articulators lay long vowel 4B 1650 Hz 1132 Hz 12591 Hz 3220 Hz mia pe J User spectrum TDS spectrum P P Z Spectrum IV Formants model s ie IV Magnitude F I Phase A Spectrum display 254 7 d El 5 0 kHz EI 0 1000 2000 3000 4000 Hz Figure 3 Vocal tract page There is the possibility to show and edit virtual EMA sensors attached to the vocal tract surfaces as used for electromagnetic articulography When EMA sensors are shown they appear as red dots in the image Any changes to the configuration of EMA sensors made via the button Edit EMA points will be saved in a configuration file when you close the program and reloaded when you start the program
11. synthesized signal Your screen should look similar to Fig 15 now Press the play buttons next to the oscillograms of the original and the synthetic signals to compare them perceptually Compare the spectrograms of the original and the synthetic signal You can change the height of the spectrograms by dragging the splitter control right above the scrollbar for the time which separates the upper and lower displays There are a few other examples of gestural scores coming with VTL 2 1 but without the master signal as in the above case Example 2 is the short German sentence Lea und Doreen m gen Bananen Lea and Doreen like bananas example 3 is the German word Vorlesung lecture and example 4 demonstrates the synthesis of five voiceless fricatives in the nonsense utterance afasafagaxa Example 5 is the sentence Hallo wie geht es dir Hello how are you 7 3 Create and save a new vocal tract shape 1 Select the vocal tract page by clicking on the corresponding tab below the toolbar 2 Open the vocal tract shapes dialog and the vocal tract dialog with the buttons Vocal tract shapes and Show vocal tract in the control panel 21 7 4 Select a shape from the list for example a with a double click on the item Drag around some of the yellow control points in the vocal tract dialog This changes the cor responding vocal tract parameters Observe the changes in the a
12. u tt alveolar fri s tt T a Min area tongue body MA1 0 05 tt alveolar fric u 3 m tt postalveol r c a Min area tongue tip MA2 0 05 tt postalveolar fri 3 etaa Min area teeth ips MA3 0 05 Move up Move down Add Replace Delete Rename Figure 8 Vocal tract shapes dialog Given the defined vocal tract shapes you can analyze the transitions from consonants to vowels in CV syllables when you call the CV transition dialog with the button CV transition dialog in the control panel Fig 9 Select the consonant from a drop down list and the vowel as either one of the vowels in the shape list fixed vowel or a vowel that results from bilinear interpolation between the predefined shapes for a i and u For the latter use the sliders to adjust the coefficients a and b Vary the slider Transition pos to generate vocal tract shapes at different positions along the CV transition The vocal tract display and all the displays on the vocal tract page are updated immediately when you change the position If your consonant is a stop you can press the button Find release position to find the position along the CV transition at which the cross sectional area after the oral release equals the number given in the edit box This allows for example to estimate formant frequencies right after the release at voice onset in plosive vowel transition which will be displayed in the spectrum dis
13. vocal tract which is drawn using OpenGL might not be displayed properly To run VTL on other platforms than Windows e g Linux or Mac we recommend to use a virtual machine for example VirtualBox www virtualbox org Note that VTL comes with no warranty of any kind The whole software is under continual development and may undergo substantial changes in future versions 1 3 Download Installation Registration The program is free of charge and available for download as a ZIP file from It needs no special installation Simply unzip the downloaded file into a folder of your choice and start the program by calling VocalTractLab2 exe The ZIP archive contains a couple of other data files most of which are example files The only essential file next to the executable is JD2 speaker which is an XML file that defines the default speaker see Sec A T and is loaded automatically when VTL is started After downloading the program it has a limited functionality To get access to all functions please register your copy of VTL Just write a short email containing your name city and institution tolpeter birkholz vocaltract lab de with the request to register free of charge In response you will receive a personalized registration file activating all the functions described in this manual Note that the response may take a few days as it is done manually If you have a registration file for a previous version of the software you can a
14. GG and Main lt gt Extra swap the signals in the corresponding tracks Use the scrollbar between the oscillogram and the spectrogram to scroll through the tracks Alternatively use Ctrl e or Ctrl gt To select a time range of the signals right click in one of the oscillogram displays and select to set the beginning or the end of the range in the context menu The time range selection is used to define the signal parts to be saved as WAV files or to be played with the Play selection button in the toolbar You can also clear the selected part of the signal with the menu item Edit Set audio selection to zero When you left click in the spectrogram or oscillogram displays you move a time mark cursor to the corresponding time in the signal The short time Fourier transform or a different transform depending on the settings in the Analysis settings dialog in Fig 2 at the time of the mark is displayed in the bottom part on the page To change the relative height of the four displays on the page you can drag the splitter controls right above and below the spectrogram window To record a microphone signal press the red button in the toolbar or press The signal will always be recorded to the main track To play back the signal in the main track press the green arrow in the tollbar or press ctrl P The green arrow between the two vertical lines or the key combination Cin plays the main trac
15. RMB right mouse button 3 Basic signal analysis Signal page The signal page provides displays and functions for the basic analysis of audio signals VTL has three tracks to store sampled signals They are called Main track EGG track and Extra track In most cases all audio analysis and synthesis takes place on the main track The EGG track is meant to store the Electroglottogram EGG signal corresponding to a speech signal in the main track Finally the extra track can store a third signal for different purposes For the copy synthesis of speech based on gestural scores see Sec 6 the extra track must contain the original master speech signal Each track represents a buffer of 60 s length with a sampling rate of 22050 Hz and a quantization of 16 bit The menu File has five items to load and save sampled signals e Load WAV EGG stereo loads a stereo WAV file The left channel is stored in the main track and the right channel is assumed to represent the corresponding EGG signal and is stored in the EGG track e Save WAV EGG stereo saves the signal in the main track and the EGG track in the selected time range see below to a stereo WAV file e Load WAV loads a mono WAV file to a track of your choice e Save WAV saves the signal in the selected time range from a track of your choice to a mono WAV file e Save WAV as TXT saves the signal in the selected time range from a track
16. Seminar on Speech Production ISSP 06 Ubatuba Brazil pp 493 500 Birkholz Peter Bernd J Kroger and Christiane Neuschaefer Rube 201 1a Articulatory synthesis of words in six voice qualities using a modified two mass model of the vocal folds In First Interna tional Workshop on Performative Speech and Singing Synthesis p3s 2011 Vancouver BC Canada 2011b Model based reproduction of articulatory trajectories for consonant vowel sequences In IEEE Transactions on Audio Speech and Language Processing 19 5 pp 1422 1433 2011c Synthesis of breathy normal and pressed phonation using a two mass model with a trian gular glottis In Interspeech 2011 Florence Italy pp 2681 2684 Birkholz Peter and Christiane Neuschaefer Rube 2012 A system for the comparison of glottal source models for articulatory speech synthesis In Sth International Conference on Voice Physiology and Biomechanics Erlangen Germany Browman Catherine P and Louis Goldstein 1992 Articulatory Phonology An Overview In Pho netica 49 pp 155 180 Fant Gunnar 1959 Acoustic analysis and synthesis of speech with applications to Swedish Ericsson Stockholm 26 Fant Gunnar Johan Liljencrants and Qi guang Lin 1985 A four parameter model of glottal flow In STL QPSR 4 pp 1 13 Ishizaka Kenzo and James L Flanagan 1972 Synthesis of Voiced Sounds From a Two Mass Model of the Vocal Co
17. VocalTractLab 2 1 User Manual Contents 1 Introduction 1 1 Purpose 1 2 Requirements Known issues Peter Birkholz September 24 2013 1 3 Download Installation Registration o oo aaa ee 2 1 Computational models 2 2 Graphical user interface GUD 2 2 ee ee 3 Basic signal analysis Signal page 4 Vocal tract model analysis Vocal tract page 5 Time domain simulation of vocal tract acoustics Time domain simulation page 6 Gestural score Gestural score page ras ones oven 7 Some typical uses 7 1 Analysis of the glottal flow for different supraglottal loads 17 2 Comparison of an utterance created by copy synthesis with its master signal 7 3 Create and save a new vocal tract shape 0 0 000000 ee 7 4 Fitting the vocal tract shape to contours in an image 200 8 Miscellaneous 8 1 Tube synthesis A File formats A l Speaker file speaker A 2 Gestural score file ges A 3 Segment sequence file seg B Changes since VTL 2 0 References 13 17 18 18 19 20 20 20 21 21 22 22 22 23 23 24 24 25 26 1 Introduction 1 1 Purpose VocalTractLab VTL is an articulatory speech synthesizer and a tool to visualize and explore the mecha nism of speech production with regard to articulation acoustics and control It is developed by Dr Peter Birkholz along with his resea
18. again so these settings will not get lost between sessions Note that the sensors can only be placed in the mid sagittal plane When you created a gestural score to define the articulatory movements for an utterance see Sec 6 the corresponding trajectories of the EMA sensors can be exported to a text file with the menu item Export EMA trajectories from gestural score Furthermore you can load an image from a GIF file with the button Load background image to be displayed in the background of the vocal tract model When you load an image with mid sagittal contours of a vocal tract you can try to adapt the articulation of the model to that shown in the image see Sec 7 4 When the box Background image editing is checked you can pan and zoom the background image by dragging with the LMB and RMB in the vocal tract display The area function corresponding to the current vocal tract shape is shown in the top left display of the vocal tract page In this image the area functions of the sinus piriformis and the nasal cavity are shown too however the sinus piriformis is currently not considered in the acoustic simulation of the vocal system by default Several display options can be found directly below the display In VTL 2 1 there is the new option to show the articulators associated with different parts of the tube This displays the articulators that confine the vocal tract at the anterior inferior side When you select thi
19. and the mapping to articulatory trajectories is the subject of ongoing research The basic principles underlying the current implementation are discussed in Birkholz and Birkholz Kr ger and Neuschaefer Rube 2011b In VTL 2 1 5th order critically damped linear systems are used to model the dynamics of articulators 2 2 P Pix Fase pe Vocal tract Time domain simulation Gestural score tabs lV Main track View range 31 ms Mark pos 0 3598 s Analysis settings ege track Detailed oscillogram Graphical user interface GUI e VocalTractLab2 Registered for Max Mustermann l e Record Play Play selection Clear nerea Era rak Calculate FO allip E Calc voice quality Control panel 3 Eje E Main lt gt EGG Overview oscillogram Main lt gt Extra tJ 2 Spectrogram of 5000 Hz 0 209 s PE N iew range 0 508 s hr ay vA 0 717 s dope ee pale Time mark ewe Kr ial pot Heo ULE UL tit N REN Eos diul erete i I Show text i i f af im iat O Hz aes ll hele aa Figure 1 Signal page Fig 1 shows the layout of the graphical user interface of VTL From top to bottom it consists of a title bar a menu bar a toolbar and the main window region The title bar shows for whom the copy of the program is registered The menu bar has the six menus File Edit Export Synthesis models Tu
20. and the vocal tract this option can be used to synthesize for example static vowels and voiced and voiceless fricatives e Gestural model is used for the synthesis of the utterance defined by the gestural score Sec 6 With this option the area function and the state of the glottis are derived from the gestural score and may change over time When this option is selected the scrollbar at the bottom of this page is enabled and allows scrolling through the utterance in time The corresponding changes in the area function over time are shown in the area function display To start the synthesis of the selected type press the button Start synthesis Continue synthesis in the control panel When the checkbox Show animation is checked the temporal change of the selected variable can be observed in the displays The speed of the animation can be changed between 1 and 100 15 Glottis models z4 T Save glottis signals during synthesis File name Use selected model for synthesis Titze Two mass model Triangular glottis Four mass model Derived values Lower rel displacement 0 01mm Upper rel displacement 0 01mm Lower abs displacement 0 14mm Upper abs displacement 0 14mm Cord length 11 44mm Lower thickness 2 73 mm Upper thickness 0 68 mm Lower area 1 73 mm 2 Upper area 1 79 mm 2 Tension Q 0 77 Contact area 0 00 mm 2 Control parameters fo 4 gt 99 74Hz Subglottal pressure lt 1000 00Pa Low
21. as VocalTractLab2 exe and is loaded automatically when the program is started Figure 17 DTD tree of the speaker file The document type definition DTD tree in shows the structure of the speaker file The root element speaker has two child elements vocal_tract_model and glottis_models The vocal_tract_model element defines the supraglottal part of the vocal tract It has the child elements anatomy and shapes The anatomy element defines the shape of the rigid parts of the vocal tract e g jaw and palate properties of the deformable structures e g velum and lips and the list of parameters that control the articulation The shapes element contains a list of shape elements Each shape defines a vocal tract target configuration for a phone in terms of vocal tract parameter values These are the shapes that are used when gestural scores are transformed into actual trajectories of vocal tract parameters The element glottis_models defines the properties of one or more vocal fold models which can be 23 used interchangeably to generate the glottal excitation of the vocal tract model in time domain simula tions of the acoustics Each of the vocal fold models is defined by an element glottis_model which in turn contains a list of glottal shapes shapes control parameters cont rol_params and static pa rameters static_params The static parameters define the speaker specific properties of the model i e the parameters that d
22. au a1 and oy there is one initial and one final shape of the vocal tract in the list Shapes for consonantal constrictions and closures are sorted by the primary articulator They start with ll for the lower lip with tt for the tongue tip and with tb for the tongue body The following part of the name indicates the place of articulation e g labial dental alveolar followed by nas for a full closure with a lowered velum nasals stop for a full closure with a raised velum plosives fric for a critical constriction fricatives and lat for a lateral constriction laterals This naming scheme is later used in gestural scores to distinguish consonantal gestures with respect to the primary articulator Note that there are three variants for consonant shapes one for the realization in each of the vowel contexts aCa iCi and uCu The context vowel is given in brackets at the end of the shape names These three variants are the foundation of the coarticulation model described in Birkholz 2013 The buttons at the bottom of the dialog allow to re sort the list and to add replace delete rename and select items When you click Add or Replace the current vocal tract configuration shown in the vocal tract dialog is added to the list or taken to replace an existing item Click Delete or Rename to delete or rename a selected item in the list The button Select tak
23. be synthesis and Help The items of these menus will be referred to in the subsequent sections The toolbar has four tools to record a sound from a microphone to play the main audio track to play a selected part of the main audio track and to clear certain user data which can be selected from a pull down menu The main window region shows one of four pages that can be selected with the page tabs below the toolbar Each page is dedicated to a certain aspect of the program The signal page which is shown in is meant for the acoustic analysis of speech signals The vocal tract page is meant for the analysis of the vocal tract model and the synthesis of vowels in the frequency domain The time domain simulation page is meant for the analysis of vocal tract acoustics in the time domain The gestural score page is meant for the manual creation of gestural scores and the copy synthesis of natural utterances The following four sections explain the basic functionality of the pages In general each of the pages consists of a main region that contains different displays and a control panel at the left side In addition to the main window many functions of VTL are controlled with individual dialogs that can be displayed or hidden on demand All the dialogs associated with the implemented synthesis models can be called 4 from the menu Synthesis models The following abbreviations will be subsequently used LMB left mouse button
24. del of the vocal tract you can specify an arbitrary vocal tract transfer function in terms of a pole zero plot This function is mainly meant for educational experiments Generally a pole zero plot displays the poles and zeros of a rational transfer function in the complex plane In the case of the vocal tract system the poles correspond to formants and the zeros to anti formants Therefore the pole zero plot allows you to define a transfer function by means of formant and anti formant frequencies and bandwidths To open the pole zero dialog shown in Fig 12 click the button Edit pole zero plot in the control panel The left side shows the location of the poles crosses and zeros circles in the bandwidth frequency plane To change the location of a pole or zero drag it with the LMB Right click in the display to call a context menu to add new or delete existing poles and zeros Press the button Enter poles to enter numerical values for formant frequencies When the check box P Z spectrum next to the spectrum display on the vocal tract page is checked the transfer function corresponding to the current pole zero plot is shown In reality a vocal tract transfer function has an infinite number of poles When you want to approximate the effect of the higher poles the ones above the highest pole that you have placed in the plot according to Fant then check the box Higher pole correction in the pole zero dialog recommended To pla
25. e Fig 13 shows the time domain simulation page which was designed for the analysis of vocal tract acous tics during the time domain simulation Here you can analyze how the pressure and volume velocity flow distribution in the vocal tract changes over time during speech production There are three displays on the page The area function display at the bottom shows the discrete area function of the vocal system Each slice of the area function represents a short section of the vocal system tube The glottis represented by two very small tube sections is about in the middle of the display To the left of the glottis the trachea is represented by 23 tube sections The tube sections of the actual vocal tract are shown at the right of the glottis At the velo pharyngeal port the nasal cavity is coupled to the vocal tract tube Its area function is flipped upside down The paranasal sinuses are modeled as Helmholtz resonators and displayed by four circles The colors of the tube sections indicate the magnitude of the selected acoustic variable at the current time of the simulation In the top right 13 Pole zero plot of a vocal tract transfer function Ea Hz Frequency Bandwidth in Hz Poles 5500 x iya 1500 70 2500 90 P000 3500 110 4500 130 45004 Poles 5500 150 formants 4000 ro 5 2 3500 x AR 4000 200 S 3000 Zero k anti formant 2500 x 2000 1500 V Higher pole correction Enter poles Enter zer
26. e in the control panel or from the menu Synthesis models LF glottal flow model With a left click in the pulse shape display you can toggle between the time function of the volume velocity and its derivative with respect to time The vowels are synthesized by the convolution of the first derivative of the glottal flow pulse sequence with the inverse Fourier Transform i e the impulse response of the vocal tract transfer function The vocal tract model comes with a number of predefined shapes for German vowels and typical conso nantal constrictions at different places of articulation These shapes are used as articulatory targets when gestural scores are transformed into time functions of vocal tract parameters The shapes are saved in the speaker file and loaded automatically when the program is started They are managed in the vocal tract shapes dialog shown in Fig 8 which is called with the button Vocal tract shapes in the control panel or from the menu Synthesis models Vocal tract shapes At the left side of the dialog the names of the vocal tract shapes are listed When a shape is selected in the list the corresponding vocal tract parameter values are shown at the right side How the vocal tract shapes were obtained is described in Birkholz 2013 The names chosen for the vocal tract shapes of vowels are the corresponding SAMPA symbols Speech Assessment Methods Phonetic Alphabet For each of the three diphthongs
27. e structure of the file for the word Banane bananoa name duration_s 0 137719 name b duration_s 0 013566 start_of_syllable 1 start_of_word 1 word_orthographic Banane word_canonic banan name a duration_s 0 072357 name n duration_s 0 114149 start_of_syllable name a duration_s 0 212388 name n duration_s 0 068383 start_of_syllable 1 name duration_s 0 195274 1 The first text line defines the length of a pause at the beginning of the utterance Each following line defines one segment in terms of its name e g SAMPA symbol and duration When a segment is the first segment of a syllable a word a phrase or a sentence this can be indicated by additional attributes in the same text line where the segment is defined For example start_of_word 1 means that the current segment is the first segment of a new word The following list shows all possible attributes for a segment e name defines the name of the segment e duration_s defines the duration of the segment in seconds e start_of_syllable set to 1 marks the start of a new syllable 24 e word_accent set to 1 places the stress in the word on the current syllable Other numbers can be used to indicate different levels of stress e phrase_accent set to 1 places the stress on the current word within the phrase e start_of_word set to 1 marks the start of a new word e word_orthographic defines the orthographic for
28. encrants Fant model for the glottal flow Fant Liljencrants and Lin 1985 with the impulse response of the vocal tract transfer function calculated in the frequency domain similar to the method by Sondhi and Schroeter 1987 The other one simulates the acoustic wave motion entirely in the time domain based on a finite difference scheme in combination with a model of the vocal folds attached to the transmission line model of the vocal tract Currently four vocal fold models are implemented the geometric model by Titze 1989 the classic two mass model by Ishizaka and Flanagan 1972 a modified two mass model by Birkholz Kroger and Neuschaefer Rube and Birkholz Kroger and Neuschaefer Rube 201 1a and a test version of an asymmetrical vocal fold model unpublished These models can be individually parameterized and exchanged for the acoustic simulation as described by Birkholz and Neuschaefer Rube 2012 Connected utterances based on gestural scores are always simulated in the time domain because this method can better account for dynamic and transient acoustic effects For the simulation of glottal and supraglottal noise sources a noise source model based on the aerodynamic acoustic relations by Stevens is used To better account for differences in noise source properties depending on the constriction position the concept of the enhanced area function to represent the vocal tract shape was recently implemented in VTL Birkholz submitted
29. er rest displacement lt gt 0 13mm Upper rest displacement lt gt 0 13mm Arytenoid area 4 gt 0 00 mm 2 Damping factor 4 gt 100 Shape modal x Save as selected l Save as existing Save as new Remove Static parameters Cord length 5 0 20 0 mm 13 000 mm L rest thickness 1 0 5 0 mm 2 400 mm a Figure 14 Glottis dialog VTL has implemented different models of the vocal folds that can be used for acoustic synthesis in the time domain These models are managed in the glottis dialog shown in which can be called with the button Glottis dialog in the control panel or from the menu Synthesis models Vocal fold models Currently four vocal fold models are implemented the geometric model by Titze 1989 the classic two mass model by Ishizaka and Flanagan 1972 the modified two mass model by Birkholz Kr ger and Neuschaefer Rube 201 1c and a four mass model The four mass model is not intended to be used yet Each model is managed on a separate dialog page that can be shown by clicking on the corresponding tab The model that is currently used for simulations is marked with a star and can be changed when the button Use selected model for synthesis is pressed Each model has a specific set of static parameters and control parameters The static parameters define speaker specific shape independent properties of the model e g the rest length and the rest mass of the vocal folds On the
30. er utterances with a gestural score because you can use the acoustic landmarks in the master signal for orientation with respect to the coordination and timing of gestures Therefore you can display the synthetic speech signal on the main track and the master signal which must be on the extra track as oscillograms and spectrograms below each other in the top right panel of this page In addition you can show the segment sequence of the master signal above the signals You can manually create edit a segment sequence and load or save it with the menu items File gt Load Save segment sequence You can insert and delete segments calling the context menu in the segment row and move the borders between segments by dragging them with the LMB A double click on a segment opens the segment annotation dialog shown in Fig 16 where the properties of a segment can be changed For a description of the properties please refer to Sec A 3 Which parts of the display are shown in the top right part of the page can be controlled in the button group Audio signals in the control panel The checkbox Show model FO curve allows showing the 19 Feature Value frame dP duration_s 0 1170 start_of_syllable word_accent phrase_accent pitch_target_offset_st pitch_target_slope_st_s start_of_word 9 word_orthographic 10 word_canonic 11 part_of_speech 12 start_of_phrase 13 start_of_sentence 14 sentence_type On au amp Wh E
31. es the selected item in the list as the current vocal tract shape To select a shape from the list and play the corresponding vowel press in the list of items To save any changes made to the shape list you must save the speaker file by pressing or selecting File Save speaker from the menu 10 E 3 0 Ol Speaker C Arbeit Programmierung VocalTractLab2 Version 2 aaa Ps Vibe wae le Horz hyoid pos HX 0 33 i o Vert hyoid pos HY 4 26 u E Horz jaw pos JX 0 00 2 y Jaw angle deg JA 4 30 A I Lip protrusion LP 0 10 E o Lip distance LD 0 81 U 9 Velum shape VS 1 00 Y Velic opening VO 0 10 6 al begin Wall compliance WC 0 00 al end OY begin Tongue body X TCX 0 12 OY end 7 aU begin Tongue body Y TCY 2 00 aU end araw Tongue tip X TTX 4 30 itaw u aw Tongue tip Y TTY 1 33 I4abial nas a I4abial nas i Tongue blade X TBX 3 07 I4abial nas u 4abial stop Tongue blade Y TBY 0 39 IHabial stop l4abial stop u Tongue root X TRX 2 81 IHabial fric a l4abial fric i Tongue root Y TRY 2 94 Iabial fric u tt alveolar nas a f a tt alveolar nas i Tongue side elevation 1 TS1 1 34 tt alveolar nas u i tt alveolar stop a Tongue side elevation 2 TS2 9 18 tt alveolar stop i x tt alveolar stop u Tongue side elevation 3 TS3 0 64 tt alveolarat a s C tt alveolarat i Tongue side elevation 4 T54 0 10 tt alveolartat
32. from the menu When an acoustic simulation of the type Phone full simulation is started the selected vocal fold model with the current parameter setting will be used The time functions of the control parameters and a number of derived parameters of a vocal fold model during an acoustic simulation can be saved to a text file for subsequent analysis The data are saved to a file when the checkbox Save glottis signals during synthesis in the upper part of the dialog is checked The file name is selected with the button File name The format of the file is as follows The first line defines the order of the saved parameters and each following line contains the parameter values of one time step of the simulation 44100 Hz sampling rate The data can be imported into MS Excel or other programs 6 Gestural score Gestural score page VocalTractLab2 Registered for Max Mustermann File Edit Export Synthesis models Tube synthesis Help Pix Signals Vocal tract Time domain simulation Gestural score S Selected segment Gesture I neutral absent Value 0 50 0 60 0 70 s 1 1 1 Sprachsynthese l x gt itn Phones tbuvuar fic Main oscillo Gesture slope E properties a Start time 0 1415 aii 122ms Main sp ctro a Main track oscillogram and spectrogram Time constant 15 0 ms TOTA Pay Time axis TUUTUU AUUT s PDEA A
33. hesize from tube sequence file The required format for the text file is described in the comment section of the example file In short you have to specify a sequence of states where each state is associated with a certain area function and 22 a certain state of the glottis the triangular glottis model is always used here Between successive states the state of the glottis and vocal tract is linearly interpolated during the synthesis The number in the first line of a specific state specifies the time from the previous state in milliseconds i e the duration of the linear interpolation Area functions must always be given in terms of 40 tube sections each with an individual length cross sectional area and articulator see Birkholz for the relevance of the articulator If you want to insert a state into a tube sequence text file that represents the current vocal tract shape and vocal fold shape in VTL you can copy the corresponding lines into the clipboard with the menu item Tube synthesis copy model states to clipboard and then paste them into your tube sequence text file A File formats A 1 Speaker file speaker The speaker file is an XML file that defines a model speaker The definition comprises the anatomy of the speaker the vocal tract shapes used to produce individual phones as well as model properties for the glottal excitation The default speaker file is JD2 speaker It must be located in the same folder
34. inuses Piriform fossa J Static pressure drops Lumped elements in T sections Figure 6 Options for the acoustic simulation in the frequency domain The vocal tract transfer function characterizes the acoustic properties of the vocal tract tube between the glottis and the lips However there are different options for the calculation of the transfer function from a given area function The options mainly regard the considered loss mechanisms They can be changed in the dialog shown in Fig 6 which is called with the button Acoustics frequency domain in the control panel The options are described in detail in Birkholz 2005 If you are not sure what an option means just leave it at the default value With the buttons Play short vowel and Play long vowel in the control panel the vocal tract model is excited with a sequence of glottal flow pulses to synthesize a short or long vowel in the main track the previous signal in the main track will be overwritten The model for the glottal flow pulses is the LF model Fant Liljencrants and Lin 1985 You can change the parameters of the model in the LF glottal 9 LF glottal flow pulse Parameters FO 4 120 00 Hz Open quotient 4 g 0 50 Shape quotient lt 4 gt 3 00 Spectral tilt 4 0 02 Pulse shape Figure 7 LF glottal flow pulse dialog flow pulse dialog shown in Fig 7 which is called with the button LF glottal flow puls
35. k signal in the selected time range This signal part is played in a loop Analysis settings r a Spectrum Spectrogram Formants GCIs FO Voice gt Window length _ gt 6 0 ms 132 pt FFT length al a gt 23 2 ms 512 pt Dynamic range al i gt 120 dB View range al _ 5000 Hz A gaussian window is used Figure 2 Dialog for analysis settings In the control panel of the page are the following buttons e Analysis settings opens the dialog shown in Fig 2 This dialog allows setting the parameters used for the calculation of the spectrum in the bottom display for the spectrogram and the fundamental frequency FO contour Furthermore a measure for voice quality along the lax tense continuum can be calculated based on the Peak Slope Parameter by Kane and Gobl 2011 Please note that the calculation of glottal closure instants and formants is not yet implemented e Calculate FO calculates the FO contour in the track of your choice which can be displayed in the spectrogram set the corresponding check mark in the analysis settings dialog on the FO tab e Analysis results opens a non modal dialog where the FO at the time of the mark in the oscillo gram spectrogram is displayed in semitones relative to the musical note Co with 16 352 Hz and Hz Furthermore the measured value for the voice quality is shown 4 Vocal tract model analysis Vocal tract page Fig 3 shows
36. lation page with the corresponding tab below the toolbar 2 Select one of the two tube sections of the glottis between the tracheal sections and the vocal tract sections with a left click in the area function display The vertical dashed line that marks the selected section should be placed as in the area function display in The time signal of the selected section and the selected acoustic variable will be shown in the upper display during simulations 3 Select the radio button Flow in the top right corner of the page 4 Open the vocal tract shapes dialog with the menu Synthesis models Vocal tract shapes Double click on the shape a in the list to make it the current vocal tract configuration 5 Click the button Set area function in the control panel to set the area function of the vocal tract for the acoustic simulation 20 7 2 Open the glottis dialog with the menu Synthesis models vocal fold models click on the tab for the Triangular glottis model and press the button Use selected model for synthesis Then select the item modal from the drop down list for the glottal shape This puts the vocal fold model into a state for modal phonation Select the radio button Phone full simulation in the control panel and press Start synthesis You should now see the time function of the glottal flow in the upper display of the page If you didn t close the glottis dial
37. line spectrum of the glottal flow of the Liljencrants Fant model LF model see below and the radiation impedance at the mouth e Radiated pressure is the spectrum of the sound pressure that would be measured in front of the mouth when the vocal tract model would be excited by the LF model e Glottal volume velocity is the line spectrum of the glottal flow of the Liljencrants Fant model Beside the spectrum selected here the Model spectrum you can display additional spectra by check ing the corresponding checkboxes right next to the display The User spectrum is the short time Fourier transform from the signal page Note that both the model primary spectrum and the user spectrum can be exported as text files from the menu Export The TDS spectrum is the Fourier transform of the im pulse response of the vocal tract calculated using the time domain simulation This allows to compare the similarity of the acoustic simulations in the frequency domain and the time domain The P Z spectrum is the transfer function defined by a pole zero plan that can be created by the user see below Options for frequency domain synthesis Em Radiation impedance Energy losses 0 lossless V Boundary layer resistance Piston in spere Wakita Fant Heat conduction losses Piston in wall 7 Soft walls Parallel circuit or R and L Hagen Poiseuille resistance Additional options Constants Paranasal s
38. ll coincide in the background image and the model You can increase the size of the vocal tract dialog for easier adjustment When you are done uncheck the checkbox Background image editing again Now try to drag the control points of the model so that the shape of the tongue the lips and so on correspond to the background image Press the button Play long vowel in the control panel of the vocal tract page to hear the corresponding vowel When the contours coincide well you should hear an u Click the button Vocal tract shapes to open the vocal tract shapes dialog and double click on the shape u raw the parameters of this shape were manually adjusted to coincide as well as possible with the background image 8 Miscellaneous 8 1 Tube synthesis With VTL 2 1 there is now the possibility to synthesize speech from a sequence of tubes i e area functions with the menu Tube synthesis If you have a sequence of area functions from any source for example extracted from X ray or real time MRI data you can compile these data into a text file and let VTL synthesize the speech To get started you can create an example tube sequence file with the menu item Tube synthesis gt Create example tube sequence file This example file defines the transition from a to 1 i e the diphthong ar Once you created the example file it can be synthesized with the menu item Tube synthesis gt Synt
39. lso use it for the new version and don t need to register again Just copy the file registration txt into the folder with the new version 2 Overview 2 1 Computational models VTL implements various models for example models for the vocal tract the vocal folds the acoustic simulation etc This section provides references to the most important models which are a recommended read if you wish to understand them in depth The core of the synthesizer is a 3D articulatory model of the vocal tract that defines the shape of the air way between the glottis and the lips It has currently 23 degrees of freedom vocal tract parameters that control the shape and position of the model articulators The model was originally developed by Birkholz 2005 and later refined by Birkholz Jack l and Kr ger 2006 and Birkholz and Kr ger 2006 The most recent version of the model is described by Birkholz 2013 For the simulation of acoustics the area function of the vocal tract is calculated i e the variation of the cross sectional area along the center line of the model The area function is then transformed into a transmission line model of the vocal tract and can be analyzed in the frequency domain or simulated in the time domain The basic method for the analysis synthesis is described by Birkholz and Jack l and Birkholz 2005 For the synthesis of vowels there are two approaches in VTL One convolves the glottal flow waveform of the Lilj
40. m of a word e word_canonic defines the canonical transcription of a word e part_of_speech defines the part of speech of a word Any values can be used here e g verb noun or adjective e start_of_phrase set to 1 indicates the start of a new phrase e start_of_sentence set to 1 indicates the start of a new sentence e sentence_type can be used to indicate the type of sentence e g it can be set to question or statement Apart from name and duration_s the use of the attributes is optional B Changes since VTL 2 0 This is an unsorted list of the major changes since VTL 2 0 e The 3D vocal tract geometry can be exported as OBJ file e g for 3D printing e VTL can automatically adjust the vocal tract shape to match specific formant frequencies by opti mization e Triangular glottis model the damping parameter was removed and a new parameter to control the strength of aspiration noise was added e A right click on one of the control points of the vocal tract model in the vocal tract dialog shows the values of the parameters that are controlled with this point e The vocal tract model can be exported as a wire frame representation into an SVG file e Dialogs always stay on top of the main window e You can attach virtual EMA sensors as in electromagnetic articulography to different vertices of the vocal tract model and export their trajectories for utterances specified in terms of gestural scores The configuratio
41. ms of the simulation The radio buttons in the control panel on the left allow you to choose the synthesis type for the simu lation e Subglottal impedance injects a short volume velocity impulse from the glottis into the trachea records the impulse response of the volume velocity and pressure right below the glottis and calculates the complex ratio of the Fourier Transform of both signals e Supraglottal impedance injects a short volume velocity impulse from the glottis into the vocal tract records the impulse response of the volume velocity and pressure right above the glottis and calculates the complex ratio of the Fourier Transform of both signals Click the button Set area function to set the area function used for the synthesis to the one of the vocal tract model e Transfer function injects a short volume velocity impulse from the glottis into the vocal tract and records the impulse response of the volume velocity at the lips The Fourier Transform of this impulse response is the vocal tract transfer function calculated in the time domain Click the button Set area function to set the area function used for the synthesis to the one of the vocal 14 VocalTractLab2 Registered for Max Mustermann Colla jme File Edit Export Synthesis models Tube synthesis Help Pbx Signals Vocal tract Time domain simulation Gestural score cm 3 s IV Show animation 400 H Signal C Pressure Speedin
42. n of EMA sensors is saved in the file config ini e You can synthesize utterances from a sequence of tube model states i e without the vocal tract model This function is in the menu Tube synthesis e The lengths of the subglottal cavity and the nasal cavity can be specified in the anatomy part of the speaker file e Vocal tract shapes for the German fricatives were added to the shape list e The dominance model for coarticulation was replaced by the interpolation model described by Birkholz 2013 e The vocal tract model was improved The velum has now two parameters VO and VS the pa rameter LH was renamed to LD and more minor changes For details see Birkholz 2013 25 e You can easily measure a time span in the audio signals on the gestural score page using the context menu e The synthesis of fricatives has been improved based on the concept of enhanced area functions Birkholz submitted e 5th order critically damped systems instead of 6th order systems are now used to model articu latory dynamics in gestural scores e Glottal shapes for phonation were adjusted to a slightly convergent shape I found this to facilitate the onset of phonation after voiceless consonants for the self oscillating triangular glottis model e The vocal tract model and the acoustic simulation have been improved in many details Acknowledgments I thank Ingmar Steiner Phil Hoole Simon Preu Felix Burkhardt and
43. o the position of the mouse When the vocal tract dialog or the glottis dialog is shown the state of the vocal tract or the glottis at the time of the mark is displayed there You can also drag the mouse from left to right with the LMB while is pressed In this case you can observe how the vocal tract shape and the glottis shape change over time Additional important keys are simultaneously increases the length of the gestures at the time mark in all tiers by a small amount In this way you can stretch the gestural score at the time of the mark for example to increase the length of the phone at that time e Ctr Der simultaneously decreases the length of the gestures at the time mark in all tiers by a small amount In this way you can shorten the gestural score at the time of the mark for example to shorten the length of the phone at that time e Ciri lt scrolls the score to the left e Cii scrolls the score to the right To synthesize the speech signal for a gestural score click the button Synthesize in the control panel for a fast simulation uncheck the checkbox Show animation The synthesized audio signal will be stored in the main track and played automatically when the synthesis finished 6 3 Copy synthesis Without a lot of practice it is quite difficult to create gestural scores for natural sounding utterances It is somewhat easier but still not trivial to copy a natural mast
44. og you also see the oscillations of the vocal fold model Wait for the simulation to finish to hear the vowel After the simulation you can replay the vowel which is stored in the main track with Select the vowel u from the list of vocal tract shapes press Set area function and start the 6699 synthesis again Observe how the glottal pulse shape differs from that for the vowel a just because of the different supraglottal configuration Comparison of an utterance created by copy synthesis with its master signal Select the gestural score page with the corresponding tab below the toolbar Load the gestural score file Example1 Sprachsynthese ges with the menu item File Load gestural score The gestural score appears in the right bottom part of the page Uncheck the checkbox Show animation and press the button Synthesize in the control panel Wait for the synthesis to finish The oscillogram and spectrogram of the synthesized signal appear in the top right display Load the audio file Examplel Sprachsynthese orig wav to the Extra track with the menu item File Load WAV Load the segment sequence Example1 Sprachsynthese seg with the menu item File Load segment sequence Check the checkboxes Show extra track and Show segmentation in the control panel to show the segment sequence and the master audio signal along with the
45. on t change over time e g the rest length and the rest mass of the vocal folds They are analogous to the anatomic part of the supraglottal vocal tract The control parameters define the properties that are controlled during articulation e g the vocal fold tension or the degree of abduction The shapes element contains a list of shape elements each of which defines a setting of the control pa rameters e g a setting for modal phonation and a setting for voiceless excitation abducted vocal folds These shapes are analogous to the vocal tract shapes for supraglottal articulation A 2 Gestural score file ges A gestural score file is an XML file that defines a gestural score The document type definition DTD tree in Fig 18 shows the structure of the file The root element is gestural_score There are eight tiers of gestures in a gestural score each of which is represented by one gesture_sequence element Each gesture sequence comprises a set of successive gestures of the same type e g vowel gestures or velic gestures The start time of a gesture is implicitly given by the sum of durations of the previous gestures of the sequence Figure 18 DTD tree of the gestural score file A 3 Segment sequence file seg A segment sequence file is a text file used to represent the structure and metadata of a spoken utterance in terms of segments phones syllables words phrases and sentences The following example illustrates th
46. ontrol by Prom on Xu and Thipakorn 2009 Glottal shape gestures control all remaining control parameters of the vocal fold model Here the target associated with a gesture is a shape that was defined for the selected vocal fold model Sec 5 In Fig 15 for example the score starts with the glottal shape open with abducted vocal folds and then approaches a state for modal phonation Supraglottal articulation is defined by the upper five tiers of gestures Here the vowel gestures define basic diphthongal movements of the vocal tract On these movements are superimposed the constriction forming gestures of the lips the tongue tip and the tongue body in the corresponding tiers The word Sprachsynthese praxzynte zo which comes as an example with VTL 2 1 and is partly show in starts for example with a vowel gesture for a for the first syllable It is superimposed with three partly overlapping consonantal gestures for f p and R to realize the consonant cluster at the beginning of the word Each gesture in the upper four tiers is associated with a particular pre defined vocal tract shape that serves as the target for the movement These are a for the vowel tt postalveolar fric for f ll labial stop for p and tb uvular fric for r a uvular fricative is produced here instead of a trill all of which can be found in the vocal tract shapes dialog All gestures defined in the gestural score a
47. os Bandwidth Play short vowel 1000 0 50 100 150 200 250 Hz Play long vowel Figure 12 Pole zero plot corner of the page you can select the variable to visualize the sound pressure or the volume velocity flow In Fig 13 the flow is visualized The red filled circles in the area function display indicate noise sources They are automatically inserted at places where turbulent noise sources are assumed to occur based on the flow conditions in the vocal tract The thick red lines mark tube sections that form critical constrictions in the vocal tract which cause the noise sources In Fig 13 one of the constrictions is the glottis Above the area function display is the space time graph Here the selected variable is shown as a curve The black curve shows the pressure or flow in the trachea the glottis and the vocal tract and the red curve shows it in the nasal cavity and the sinus piriformis The ordinate of both displays can be scaled with the and buttons left of the displays In the area function display you can select one of the tube sections with the LMB The selected section is marked with a vertical dashed line InfFig 13 the upper glottis section is selected In the upper display of this page the time signal display the selected variable in the selected tube section is plotted as a function of time for the last 50 ms The signal shown in Fig T3 is therefore the glottal volume velocity in the last 50
48. play The button Improve formants in the control panel opens the formant optimization dialog shown in It allows the automatic adjustment of the vocal tract shape such that the formants in the transfer function approach specific frequencies given by the user An algorithm tries to find vocal tract parameter values in the vicinity of the current vocal tract shape that change the formant frequencies towards the given values This may be interesting for example when you want to adapt the formants of a vowel to the realization of that vowel in a different language or dialect Enter either two three or four formant frequencies at the top of the dialog and use the checkboxes below to define which of the vocal tract parameters are not supposed to be varied during the optimization You can furthermore set an upper limit of the displacement of the vocal tract contour during optimization and a minimal cross sectional area The latter should typically not drop below 25 cm for vowels because the pressure drop would become too high and possibly prevent vocal fold oscillations otherwise Press Optimize vowel to start 11 CV transition dialog Ex Consonant SREE v Update lists Fixed vowel a v T 0 00 0 00 L 0 00 0 00 Interpolated vowel a 4 gt 0 000 b 4 _ gt 0 000 Transition pos _4 gt 0 0 Find release position atareaof 20 0 mm 2 Cancel Figure 9 Consonant vowel transition dialog the optimization Yo
49. ract transfer function corresponding to the current shape of the vocal tract This spectrum is calculated in the frequency domain based on the transmission line model of the vocal tract cf Sec 2 1 The formant frequencies are automatically determined and marked by vertical dashed lines Note that the formant determination is not reliable when there are zeros anti formants in the transfer function for example when the velo pharyngeal port is open A left click in the spectrum display opens the dialog in Fig 5 Here you can select the kind of model spectrum to display e Transfer function U U is the volume velocity transfer function between the glottis and the lips e Transfer function P U is the complex ratio of the radiated sound pressure and the volume velocity at the glottis e Input impedance of the vocal tract is the input impedance seen from the glottis e Input impedance of the subglottal system is the input impedance of the subglottal system seen from the glottis Options for the primary spectrum x Spectrum type Transfer function U U Transfer function P U Input impedance of the vocal tract Input impedance of the subglottal system Glottal volume velocity radiation impedance Radiated pressure Glottal volume velocity Options Show transfer function harmonics Figure 5 Options for the model spectrum to display e Glottal volume velocity radiation impedance is the product of the
50. rch on articulatory speech synthesis With VTL you can for example e analyze the relationship between articulation and acoustics of the vocal tract e synthesize vowels and consonants from arbitrary vocal tract shapes with different models for glot tal excitation e synthesize vowels from an arbitrary set of formants and anti formants e synthesize connected speech utterances based on gestural scores e analyze the time varying distribution of pressure and volume velocity within the vocal tract during the synthesis of speech However VTL is not a text to speech system At the moment connected utterances can only be synthe sized based on gestural scores as described in Sec 6 1 2 Requirements Known issues VTL is currently developed for the Windows platform It was tested with Windows 7 and Windows XP but should also run under Windows Vista and 8 The minimum screen resolution is 1024x768 but higher resolutions are preferable As the simulations are computationally intensive a fast computer is recommended For these reasons tablet computers and netbooks are generally not suited to work with VTL There was no extensive testing of the software but the parts of the program made available are considered to be relatively stable Please feel free to report any bugs to Currently there is the following issue With Windows Vista and 7 you should use the Aero design for the desktop this is the default Otherwise the model of the
51. rds In The Bell System Technical Journal 51 6 pp 1233 1268 Kane J and C Gobl 2011 Identifying regions of non modal phonation using features of the wavelet transform In Interspeech 2011 Florence Italy pp 177 180 Prom on Santhitam Yi Xu and Bundit Thipakorn 2009 Modeling Tone and Intonation in Mandarin and English as a Process of Target Approximation In Journal of the Acoustical Society of America 125 1 pp 405 424 Sondhi Man Mohan and Juergen Schroeter 1987 A hybrid time frequency domain articulatory speech synthesizer In IEEE Transactions on Acoustics Speech and Signal Processing ASSP 35 7 pp 955 967 Stevens Kenneth N 1998 Acoustic Phonetics The MIT Press Cambridge Massachusetts Titze Ingo R 1989 A Four Parameter Model of the Glottis and Vocal Fold Contact Area In Speech Communication 8 pp 191 201 27
52. re either active gestures or neutral gestures An active gesture is painted as a gray box and specifies a certain target for the vocal tract model or vocal fold model parameters A neutral gesture is painted as a white box and represents the movement towards a neutral or underlying target For example in the tier of tongue body gestures the active and selected gesture for the uvular fricative is followed by a neutral gesture that simply represents the movement back to the underlying vocalic target for a In each tier the time function of a representative vocal tract vocal fold parameter is drawn These curves are just meant to illustrate the effect of the gestures on the movements The curve in the tier of tongue tip gestures shows for example the vertical position of the tongue tip Please consider that the details of gestural control in VTL are under active development and may undergo major changes in the future 6 2 Editing a gestural score Click the LMB on a gesture to select it The selected gesture is painted in yellow The properties of the selected gesture can be controlled in the upper part of the control panel Each gesture has a duration 18 a time constant and a value The duration defines the length of the gesture in the gestural score and the time constant specifies how quickly the participating articulators reach the target associated with the gesture A high time constant results in a slow approach and a low time
53. rea function and the vocal tract transfer function Press the button Play long vowel in the control panel to hear the sound corre sponding to the current articulation Press the button Add in the vocal tract shapes dialog to add the current vocal tract configuration to the list of shapes Type test as the name for the new shape and click OR Click on the menu item File gt Save speaker to save the speaker file JD2 speaker in order to permanently save the new shape When you now close VTL and start it again the speaker file containing the extended list of shapes is automatically loaded Fitting the vocal tract shape to contours in an image Select the vocal tract page with the corresponding tab below the toolbar Click the button Show vocal tract in the control panel to show the vocal tract dialog Click the button Load background image in the dialog and load the file vowel u outline gif This image shows the outline of the vocal tract for the vowel u obtained from MRI data The image can be seen behind the 3D vocal tract model Click the radio button 2D in the vocal tract dialog to show the mid sagittal outline of the vocal tract model Check the checkbox Background image editing at the bottom of the dialog Now drag the background image with the LMB and scale it by dragging the RMB such that the outline of the hard palate and the rear pharyngeal wa
54. s display option the tongue and the lower lip regions are shown in orange and the lower incisor region is ivory colored The remaining parts of the vocal tract are shown in gray This association between tube sections and articulators was introduced to improve the modeling of noise sources as described in Birkholz submitted Tongue side elev mm Minimal area cm 2 a jof336 alj _ 0 05 af il 27 _ 0 05 af J 638 all _ gt 0 05 al lil _ 1 00 V Automatic TRX TRY calc Cx CD 3 C Wire frame M Both sides IV Show control points Show cut vectors Show center line V smooth J Show EMA points Edit EMA points Load background image No image loaded IV Show background image Background image editing Figure 4 Vocal tract dialog The display right next to the area function shows a cross section through the 3D vocal tract model from which the area grey region was calculated at a certain position along the center line The position of this cross section is marked by the vertical dashed line in the area function display and is also shown in the vocal tract display when the box Show center line is checked in the dialog The position along the center line can be changed by dragging the corresponding control point or the vertical dashed line in the area function display The display in the bottom part of the vocal tract page shows one or more spectra The spectrum shown by default is the vocal t
55. the vocal tract page which was designed for the articulatory acoustic analysis of the vocal tract The actual 3D vocal tract model is shown in a separate dialog Fig 4p This dialog is opened automatically when the program is started If it has been closed it can be shown again by selecting the menu item Synthesis models Vocal tract model or by pressing the button Show vocal tract in the control panel The parameters of the model can be changed by dragging the yellow control points with the LMB When no control point is selected dragging the LMB in the vocal tract display turns the model around two axes When the mouse is dragged with the RMB the model can be zoomed in and out Seven vocal tract parameters which cannot be controlled with control points are adjusted using the scrollbars below the vocal tract display tongue side elevation and minimal area parameters When the box Automatic TRX TRY calc is checked the vocal tract parameters TRX and TRY which specify the shape of the tongue root are calculated by the software based on the other parameter values Because the shape of the tongue root is well predictable for speaker JD 1 this is the recommended option If the box is unchecked a control point will appear for the tongue root Display options for the vocal tract model can be found at the bottom of the dialog VocalTractLab2 Registered for Max Mustermann kodak File Edit Export Synthesis models Tube
56. u can also optimize the vocal tract shape of plosives with respect to the formant onset frequencies after their release towards a given vowel in the bottom part of the dialog Besides the level of the elementary vocal tract parameters which are shown in the vocal tract shapes dialog VTL provides another higher level of phonetic parameters to control the vocal tract shape This level of parameters is mainly meant for educational experiments Click the button Phonetic parame ters in the control panel or select Synthesis models Phonetic parameters from the menu to open the phonetic parameter dialog shown in There are seven parameters which can be changed by scrollbars The parameters Tongue height and Tongue frontness specify the tongue shape for a vowel The actual shape is calculated by bilinear interpolation between the predefined shapes for the comer vowels a i and u The degrees of lip rounding and velum lowering are separately specified by the parameters Lip rounding and Velum position Finally the parameters Bilabial constriction degree Apico alveolar constriction degree and Dorso velar constriction degree can be used to su perimpose consonantal constrictions of varying degrees on the vocalic configuration When a parameter is changed the corresponding changes in the vocal tract shape the area function and the transfer function are immediately displayed Independently of the mo
57. y the short or long vowel sound corresponding to the current pole zero plot press the button Play short vowel or Play long vowel The audio signal for the vowel is synthesized using the LF glottal pulse model with its current parameter settings and it is stored in the main audio track 12 Formant optimization First two or three formant frequencies separated by spaces 500 1500 2500 Please check the parameters that are fixed Tux Ca fEx Fa Fe Fw Livs Live m iitec Li tev Lita L try tex tey TRX try TSi Tts2 Ts3 154 V MA1 M MA2 V MA3 Max contour displacement 2 0 mm Min cross sectional area 25 0 mm 2 Optimize vowel Release area 20 0 mm 2 for consonants Context vowel a for consonants Optimize consonant Cancel Set default values Figure 10 Formant optimization dialog Phonetic parameters x The generation of vocal tract shapes from these parameters refers to the vocal tract shapes a i u l4abial stop tt alveolar stop and tb velar stop Make sure that these shapes are in the vocal tract shape list Tongue height al J p 50 Tongue frontness a J 50 Lip rounding 4 p 30 Velum position 4 gt 0 Bilabial constriction degree 4 r 0 Apico alveolar const degree q gt 0 Dorso velar const degree 4 p 0 Figure 11 Phonetic parameters dialog 5 Time domain simulation of vocal tract acoustics Time domain simula tion pag
Download Pdf Manuals
Related Search