29.ijaest vol no 8 issue no 2 study of various kinds of speech synthesizer technologies and...

Upload: luu-minh-tien

Post on 05-Apr-2018

212 views

Category:

Documents


0 download

TRANSCRIPT

  • 7/31/2019 29.IJAEST Vol No 8 Issue No 2 Study of Various Kinds of Speech Synthesizer Technologies and Expression for Expre

    1/5

    Study of Various kinds of Speech Synthesizer Technologies and Expression For Expressive Text To

    Speech Conversion System

    Suresh Kumar Thakur Computer Science Engineering

    Rungta College Of Engineering & TechnologyBhilai , India

    [email protected]

    K.J. SataoInformation Technology

    Rungta College Of Engineering & TechnologyBhilai , India

    [email protected]

    Abstract This paper surveys various types of text to speechconversion methods. It discusses speech synthesis techniques andexpression with various parameters, with the goal of identifying

    techniques most suitable to design and implementation of expressive Text to speech Conversion system. In this manuscriptwe define the techniques to generate the speech as output bytaking the input as text. Text to Speech synthesis we can, intheory, mean any kind of method to generate of speech. Each of the synthesizers received the text and the necessarypronunciation corrections as input and had to generate speechoutput. All options of each synthesizer were set to their defaultvalues, to the settings recommended by the manufacturers. Thiswas done for the following parameters pitch level, pitchdynamics, aspiration, rate, and volume.

    Keywords-survey; expressive; parameter; speech ; synthesizer(key words)

    I. I NTRODUCTIONSpeech synthesis is the artificial production of human

    speech. A computer system used for this purpose is called aspeech synthesizer, and can be implemented in software or hardware. A text-to-speech (TTS) system converts normallanguage text into speech; other systems render symboliclinguistic representations like phonetic transcriptions intospeech

    Synthesized speech can be created by concatenating piecesof recorded speech that are stored in a database. Systems differ in the size of the stored speech units; a system that stores phones or diphones provides the largest output range, but maylack clarity. For specific usage domains, the storage of entirewords or sentences allows for high-quality output.Alternatively, a synthesizer can incorporate a model of thevocal tract and other human voice characteristics to create acompletely synthetic voice output.

    The quality of a speech synthesizer is judged by itssimilarity to the human voice and by its ability to beunderstood. An intelligible text-to-speech program allows people with visual impairments or reading disabilities to listento written works on a home computer. Many computer

    operating systems have included speech synthesizers since theearly 1980s.

    Text to Speech synthesis we can, in theory, mean any kindof synthesizers f speech. For example, it can be the process inwhich a speech decoder generates the speech signal based onthe parameters it has received through the transmission line, or it can be a procedure performed by a computer to estimatesome kind of a presentation of the speech signal given a textinput. Text-to-speech synthesis is a research field that hasreceived a lot of attention and resources during the last coupleof decades for excellent reasons. One of the most interestingideas is the fact that a workable Text To Speech system,combined with a workable speech recognition device, wouldactually be an extremely efficient method for speech coding. Itwould provide incomparable compression ratio and flexible possibilities to choose the type of speech, the fundamental

    frequency along with its range, the rhythm of speech, andseveral other effects. Furthermore, if the content of a messageneeds to be changed, it is much easier to retype the text than torecord the signal again. Of course there are also numerousspeech synthesis applications that are closer to being available.For instance, a telephone inquiry system where the informationis frequently updated can use Text to speech to deliver answersto the customers. Speech synthesizers are also important to thevisually impaired and to those who have lost their ability tospeak. Several other examples can be found in everyday life,such as listening to the messages and news instead of readingthem, and using hands-free functions through a voice interfacein a car, and so on.

    The most commonly used criteria for high-quality speechare intelligibility, naturalness and pleasantness. Since these aremultidimensional factors that depend on each other, thecomprehensive high quality is formed by the interaction of numerous details. Thus, the elimination of background noise,musical noise, mumbling, and the various pops and cracks,does not result in the ultimate quality, but the speech shouldalso be made rich in nuances and it should carry informationabout the personality of the speaker. Moreover, it would beadvisable to include in the speech some features that describethe emotional state of the speaker because this improves thenaturalness and makes the speech more lively. The quality of

    Suresh Kumar Thakur* et al. / (IJAEST) INTERNATIONAL JOURNAL OF ADVANCED ENGINEERING SCIENCES AND TECHNOLOGIESVol No. 8, Issue No. 2, 301 - 305

    ISSN: 2230-7818 @ 2011 http://www.ijaest.iserp.org. All rights Reserved. Page 301

  • 7/31/2019 29.IJAEST Vol No 8 Issue No 2 Study of Various Kinds of Speech Synthesizer Technologies and Expression for Expre

    2/5

    speech synthesizers can also be examined apart from speechquality as such, in which case we are interested in, for instance,the following facts: How much the operation of the synthesizer depends on the surrounding software or the operating system?How good is the performance of the synthesizer in the particular application for which it has been designed? Howmuch memory does it require? Given the continuouslyincreasing processing capacity, more and more complicated

    speech synthesizers can be used even on personal computers.

    Fig. 1 Text to speech Conversion process

    The overall goal of the speech synthesis researchcommunity is to create natural sounding synthetic speech. Toincrease naturalness, researchers have been interested insynthesizing emotional speech for a long time. One way

    synthesized speech benefits from emotions is by deliveringcertain content in the right emotion (e.g. good news aredelivered in a happy voice), therefore making the speech andthe content more believable. Emotions can make theinteraction with the computer more natural because the systemreacts in ways that the user expects. Emotional speechsynthesis is a step in this direction The implementation of emotions seems straightforward at first but a closer look reveals many difficulties in studying and implementingemotions. The difficulties start with the definition of emotions.Researchers agree that emotions are not as often thought of, just a subjective experience or feeling.

    II.

    TEXT TO SPEECH SYNTHESIZER TECHNOLOGIESThe most important qualities of a speech synthesis system

    are naturalness and intelligibility. Naturalness describes howclosely the output sounds like human speech, whileintelligibility is the ease with which the output is understood.The ideal speech synthesizer is both natural and intelligible.Speech synthesis systems usually try to maximize bothcharacteristics. The two primary technologies for generatingsynthetic speech waveforms are concatenative synthesis andformant synthesis. Each technology has strengths and

    weaknesses, and the intended uses of a synthesis system willtypically determine which approach is used.

    A. Concatenative synthesis

    Concatenative synthesis is based on the contaminative (or stringing together) of segments of recorded speech. Generally,concatenative synthesis produces the most natural-soundingsynthesized speech. However, differences between naturalvariations in speech and the nature of the automatedtechniques for segmenting the waveforms sometimes result inaudible glitches in the output. There are three main sub-typesof concatenative synthesis Concatenative synthesis is atechnique for synthesizing sounds by concatenating shortsamples of recorded sound (calledunits ). The duration of theunits is not strictly defined and may vary according to theimplementation, roughly in the range of 10 milliseconds up to1 second. It is used in speech synthesis and music soundsynthesis to generate user-specified sequences of sound from adatabase built from recordings of other sequences.

    Concatenative text-to-speech systems can, in theory, produce very naturally sounding synthetic speech, since theysimply join pre-recorded segments or units to form anysentence. In practice, several factors contribute for less perfectspeech output quality. For instance, the choice of the best setof pre-recorded speech units that can be used as building blocks is a difficult task. Moreover, the concatenation of unitsrecorded using different intonation or phonetic contexts may produce suboptimal results even if the set is reasonablycomplete and if some prosodic transformations are performedduring the concatenation phase. [1]

    B. Unit selection synthesis

    Unit selection synthesis uses large databases of recordedspeech. During database creation, each recorded utterance issegmented into some or all of the following: individual phones,diaphones, half-phones, syllables, morphemes, words, phrases,and sentences. Typically, the division into segments is doneusing a specially modified speech recognizer set to a "forcedalignment" mode with some manual correction afterward,using visual representations such as the waveform andspectrogram. An index of the units in the speech database isthen created based on the segmentation and acoustic parameters like the fundamental frequency (pitch), duration, position in the syllable, and neighboring phones. At runtime,the desired target utterance is created by determining the bestchain of candidate units from the database (unit selection). This process is typically achieved using a specially weighteddecision tree. Unit selection provides the greatest naturalness, because it applies only a small amount of digital signals processing (DSP) to the recorded speech. DSP often makesrecorded speech sound less natural, although some systems usea small amount of signal processing at the point of concatenation to smooth the waveform.

    The output from the best unit-selection systems is oftenindistinguishable from real human voices, especially inContexts for which the TTS system has been tuned. However,

    Suresh Kumar Thakur* et al. / (IJAEST) INTERNATIONAL JOURNAL OF ADVANCED ENGINEERING SCIENCES AND TECHNOLOGIESVol No. 8, Issue No. 2, 301 - 305

    ISSN: 2230-7818 @ 2011 http://www.ijaest.iserp.org. All rights Reserved. Page 302

  • 7/31/2019 29.IJAEST Vol No 8 Issue No 2 Study of Various Kinds of Speech Synthesizer Technologies and Expression for Expre

    3/5

    maximum naturalness typically require unit-selection speechdatabases to be very large, in some systems ranging into thegigabytes of recorded data, representing dozens of hours of speech. Also, unit selection algorithms have been known toselect segments from a place that results in less than idealsynthesis (e.g. minor words become unclear) even when a better choice exists in the database .

    The goal of unit selection is to provide a mechanismwhereby segments of prerecorded speech are selected for synthesis. These segments are provided from a database. Theymay be parameterized in some form (LPC, HNM) or other or may be digitized speech samples.[2]

    C. Diphone synthesis

    Diphone synthesis uses a minimal speech databasecontaining all the diphones (sound-to-sound transitions)occurring in a language. The number of diphones depends onthe phonotactics of the language: for example, Spanish hasabout 800 diphones and German about 2500. In diphonesynthesis, only one example of each diphone is contained in thespeech database. At runtime, the target prosody of a sentence is

    superimposed on these minimal units by means of digital signal processing techniques such as linear predictive coding,PSOLA[4] or MBROLA[5]. The quality of the resulting speechis generally worse than that of unit-selection systems, but morenatural-sounding than the output of formant synthesizers.

    Diphone synthesis suffers from the sonic glitches of concatenative synthesis and the robotic-sounding nature of formant synthesis, and has few of the advantages of either approach other than small size. As such, its use in commercialapplications is declining, although it continues to be used inresearch because there are a number of freely availablesoftware implementations.

    Diphone-based speech synthesis has been researched for many years. Some of the most successful diphone systems have been concatenative [6]. Such systems can produce veryintelligible synthetic speech, but tend not to sound completelynatural. This lack of naturalness can be attributed, at least in part, to the limited set of units from which speech is chosen(typically ~2000 diphones), coupled with the need to prosodically modify the speech signal of each diphone.[3]

    D. Domain-specific synthesis

    Domain-specific synthesis concatenates pre-recorded wordsand phrases to create complete utterances. It is used inapplications where the variety of texts the system will output islimited to a particular domain, like transit schedule

    announcements or weather reports.[6] The technology is verysimple to implement, and has been in commercial use for along time, in devices like talking clocks and calculators.

    The term text-to-speech (TTS) synthesis is used to describethe process of converting given raw text into synthetic speech.Concept-to-speech (CTS) is a term often used for speechsynthesis where the input is not text, but rather a machinegenerated message. We can think of a TTS system ascomprising two main components: text analysis and speechgeneration. The text analysis component has to resolve theambiguities inherent in written text and produce a clean

    linguistic representation of the sentence to be spoken, e.g.appropriate word stress. In CTS, the situation is very different.There is no prior input text as such; rather a natural languagegeneration (NLG) system generates some text from scratch.

    The level of naturalness of these systems can be very high because the variety of sentence types is limited, and theyclosely match the prosody and intonation of the originalrecordings. Because these systems are limited by the words and phrases in their databases, they are not general purpose and canonly synthesize the combinations of words and phrases withwhich they have been pre programmed. The blending of wordswithin naturally spoken language however can still cause problems unless the many variations are taken into account.For example, in non-chaotic dialects of English the "r" inwords like "clear" /kli/ is usually only pronounced when thefollowing word has a vowel as its first letter (e.g. "clear out" isrealized as /kli t/). Likewise in French, many finalconsonants become no longer silent if followed by a word that begins with a vowel, an effect called liaison. This alternationcannot be reproduced by a simple word concatenation system,which would require additional complexity to be context-

    sensitive. E. Formant synthesis

    Synthesis does not use human speech samples at runtime.Instead, the synthesized speech output is created using additivesynthesis and an acoustic model (physical modelingsynthesis)[7]. Parameters such as fundamental frequency,voicing, and noise levels are varied over time to create awaveform of artificial speech. This method is sometimes calledrules-based synthesis; however, many concatenative systemsalso have rules-based components. Many systems based onformant synthesis technology generate artificial, roboticsounding speech that would never be mistaken for humanspeech. However, maximum naturalness is not always the goalof a speech synthesis system, and formant synthesis systemshave advantages over concatenative systems. Formantsynthesized speech can be reliably intelligible, even at veryhigh speeds, avoiding the acoustic glitches that commonly plague concatenative systems. High-speed synthesized speechis used by the visually impaired to quickly navigate computersusing a screen reader. Formant synthesizers are usually smaller programs than concatenative systems because they do not havea database of speech samples. They can therefore be used inembedded systems, where memory and microprocessor power are especially limited. Because formant-based systems havecomplete control of all aspects of the output speech, a widevariety of prosodies and intonations can be output, conveying

    not just questions and statements, but a variety of emotions andtones of voice.In formant synthesis, the basic assumption is that the vocal

    tract transfer function can be satisfactorily modeled bysimulating formant frequencies and formant amplitudes. Thesynthesis thus consists of the artificial reconstruction of theformant characteristics to be produced. This is done by excitinga set of resonators by a voicing source or noise generator toachieve the desired speech spectrum, and by controlling theexcitation source to simulate both voicing and voice lessons.

    Suresh Kumar Thakur* et al. / (IJAEST) INTERNATIONAL JOURNAL OF ADVANCED ENGINEERING SCIENCES AND TECHNOLOGIESVol No. 8, Issue No. 2, 301 - 305

    ISSN: 2230-7818 @ 2011 http://www.ijaest.iserp.org. All rights Reserved. Page 303

  • 7/31/2019 29.IJAEST Vol No 8 Issue No 2 Study of Various Kinds of Speech Synthesizer Technologies and Expression for Expre

    4/5

    The addition of a set of anti-resonators furthermore allows thesimulation of nasal tract effects, fricatives and plosives.[4]

    The specification of about 20 or more such parameters canlead to a satisfactory restitution of the speech signal. Theadvantage of this technique is that its parameters are highlycorrelated with the production and propagation of sound in theoral tract. The main current drawback of this approach is thatautomatic techniques of specifying formant parameters are stilllargely unsatisfactory, and that consequently, the majority of parameters must still be manually optimized.[4]

    F. Articulatory synthesis

    Articulatory synthesis refers to computational techniques for synthesizing speech based on models of the human vocal tractand the articulation processes occurring there. The firstarticulatory synthesizer regularly used for laboratoryexperiments was developed at Haskins Laboratories in the mid-1970s by Philip Rubin, Tom Baer, and Paul Mermelstein. Thissynthesizer, known as ASY, was based on vocal tract modelsdeveloped at Bell Laboratories in the 1960s and 1970s by PaulMermelstein, Cecil Coker, and colleagues. Until recently,

    articulatory synthesis models have not been incorporated intocommercial speech synthesis systems. A notable exception isthe NeXT-based system originally developed and marketed byTrillium Sound Research, a spinoff company of the Universityof Calgary, where much of the original research wasconducted. Following the demise of the various incarnations of NeXT (started by Steve Jobs in the late 1980s and merged withApple Computer in 1997), the Trillium software was publishedunder the GNU General Public License, with work continuingas gnu speech. The system, first marketed in 1994, providesfull articulatory-based text-to speech conversion using awaveguide or transmission-line analog of the human oral andnasal tracts controlled by Carr's "distinctive region model".

    In this note one aspect of articulatory synthesis will beconsidered-that of solving the differential equation describingacoustic (small amplitude), one-dimensional propagation of a pressure disturbance through a lossless tube with spatiallyvarying cross section.

    III. EXPRESSION FOR DEVELOPINGEXPRESSIVETEXTTOSPEECHCONVERSIONSYSTEM

    According to this project we try to implement text to speechsystem by integrating following basic emotions which is to beexplained as below:

    A. Anger This is the emotional category where findings from both

    spontaneous and elicited material consistently report featuressuch as high mean, wide pitch range, high energy and fasttempo. In detail Fairbanks and Provonost (neutral phrase,acted speech, subjects recognized emotions, English) reportthat anger generally is characterized by high pitch and a wide pitch range. Fnagy & Magdics ( recordings of spokenHungarian, subjective analyzing,) reports that anger ischaracterized by mid pitch and a straight rigid melody. ster and Risberg (conversation, perception experiment with

    hearing impaired subjects, Swedish) reports high tempo,normal pitch range and normal pitch level.Anger is an emotion related to one's psychologicalinterpretation of having been offended, wronged or denied anda tendency to undo that by retaliation. Videbeck [7] describesanger as a normal emotion that involves a stronguncomfortable and emotional response to a perceived provocation. R. Novaco recognized three modalities of anger:cognitive (appraisals), somatic-affective (tension andagitations) and behavioral (withdrawal and antagonism). DeFoore. W 2004 describes anger as a pressure cooker; we canonly apply pressure against our anger for a certain amount of time until it explodes. Anger may have physical correlatessuch as increased heart rate, blood pressure, and levels of adrenaline and nor adrenaline[8]. Some view anger as part of the fight or flight brain response to the perceived threat of harm [8] Anger becomes the predominant feeling behaviorally, cognitively, and physiologically when a personmakes the conscious choice to take action to immediately stopthe threatening behavior of another outside force[9] TheEnglish term originally comes from the termanger of Old Norse language[10]. Anger can have many physical andmental consequences

    B. Joy, happiness, humour Generally researchers have focused on more extreme forms

    of happiness, but still, here the findings reported are quiteconsistent between researchers. However, Stibbard points outthat concerning voice quality the findings are contradictory. Indetail Skinner (, emotions induced with music, subjectsvocalised emotions through "ah", new subjects recognised theemotions), Cowan , acted speech, subjective enelysis andster & Risberg , report that happiness gives an increase in pitch and pitch range. ster and Risberg noted a slow tempo,while Fnagy & Magdics, conversations, acted speech, music,subjective analysis, Hungarian) described it as lively. Davitz, subjects rated stimuli with 14 emotive adjectives) reported anincrease in speech rate along with an increase in intensity, thisintensity also being noted by Skinner .

    C. SadnessThe equations are an exception to the prescribed

    specifications of this template. General findings describesadness as exhibiting normal or lower pitch, narrow pitch rangeand slow tempo , Davitz, Fnagy , Murray and Arnott describethe material as being "a wide range of analysis andexperimental techniques", ster and Risberg, However, there

    are contradictory findings concerning voice quality D. Fear/Anxiety

    Generally reported features are increased mean F0,increased F0 range, increased F0 Fairbanks and Pronovost(1939, neutral phrase spoken with different emotive expression by non-actors, subjects recognised emotions) reported therelatively highest pitch and the widest pitch range and thehighest pitch median. Fairbanks and Hoaglin (, neutral phrasespoken with different emotive expression by non-actors,subjects recognised emotions) noted high speech rate. Williams

    Suresh Kumar Thakur* et al. / (IJAEST) INTERNATIONAL JOURNAL OF ADVANCED ENGINEERING SCIENCES AND TECHNOLOGIESVol No. 8, Issue No. 2, 301 - 305

    ISSN: 2230-7818 @ 2011 http://www.ijaest.iserp.org. All rights Reserved. Page 304

  • 7/31/2019 29.IJAEST Vol No 8 Issue No 2 Study of Various Kinds of Speech Synthesizer Technologies and Expression for Expre

    5/5

    and Steven , acted speech, emotions induced by content inspecially written play with control clusters) reported low F0, but with occasional F0 peaks, low speech rate.

    E. Disgust For the category disgust induced data and acted data has

    shown directly contradictory results Studies using induceddata reports an increase in mean F0, while studies using acted

    data found a decrease .F. Hatred/contempt/scorn

    Fairbanks & Pronovost noted low pitch median, wide pitch range and very wide downward inflections at phraseendings. Fairbanks & Hoaglin noted lowest observed speechrate, caused by prolonged speech rate and increased pauselength. Fnagy & Magdics noted that scorn is reflected by adescending melodic line, intoned to a very low level.However, Stibbard points out, that the only parameter analysed in studies more than one is tempo, why it is difficultto evaluate the findings.

    IV. CONCLUSION ANDDISCUSSION Using current methods such as those outlined in this paper,

    it is possible to produce speech output that is of highintelligibility and reasonable naturalness, given unrestrictedinput text. However, there is still much work to be done in allareas of the problem, including: improving voice quality, andallowing for greater user control over aspects of voice quality; producing better models of intonation to allow for morenatural-sounding contours; and improving linguistic analysis sothat more accurate information on contextually appropriateword pronunciation, accenting, and phrasing can be computedautomatically. This latter are linguistic analysis is particularlycrucial: most high-quality TTS systems allow for user controlof the output speech by means of various escape sequences,which can be inserted into the input text. By use of such escapesequences, it is possible to produce highly appropriate andnatural sounding output. What are still lacking in many casesare natural-language analysis techniques that can mimic what ahuman annotator is able to do.

    This paper explore the different type of method to generatetext to speech conversion in computer technology , or the purpose of developing expressive text to speech conversionsystem we need to add the various emotion which is explainedabove in this paper. To generate emotion for adding in to the proposed system we try to control the parameter such as pitch,rate and volume. This is the basic parameter for generating

    proposed system that convert inputted text data or informationin form of speech or sound data.

    The actual experiment took place on a computer wheresubjects had to rate the synthesized sentences. The rating wasdone using a continues scale represented by a slider bar thatwas labeled angry on the left side, happy on the right side andneutral in the centre. The program recorded the input on a scalefrom 0 to 100 where 50 meant neutral, 0 meant angry, and 100meant happy. It is not correct to say that happy and angry areopposite of each other but for the purpose of this experiment itshould not had any effect on the outcome. This type of ratingscale had the advantage that it is forced choice with a continuosresponse. A continuos response was needed because thestrength of a given emotion was part of the assessment.

    R EFERENCES

    [1] Marc Schroder and Martine Grice "Expressing vocal effort inconcatenative synthesis" 15th ICPhS Barcelona pp 2589-2592.

    [2] Alistair Conkie "ROBUST UNIT SELECTION SYSTEM FOR SPEECH SYNTHESIS".

    [3] Thomas Styger and Eric Keller "Formant Synthesis" Styger, T., &Keller, E. (1994). Formant synthesis.In E. Keller (ed.), Fundamentals of Speech Synthesis and Speech Recognition: Basic Concepts, State of theArt, and Future Challenges (pp. 109-128)..

    [4] David hlin and Rolf Carlson "Data-driven formant synthesis"Proceedings, FONETIK 2004, Dept. of Linguistics, StockholmUniversity.

    [5] P A TAYLOR "Concept-to-Speech Synthesis by Phonological StructureMatching".

    [6] Videbeck, Sheila L. (2006.). Psychiatric Mental Health Nursing (3rded.). Lippincott Williams & Wilkins.

    [7] "Anger definition". Medicine.net. Retrieved 2008-04-05.[8] Harris, W., Schoenfeld, C. D., Gwynne, P. W., Weissler, A.

    M.,Circulatory and humoral responses to fear and anger, ThePhysiologist, 1964, 7, 155.

    [9] Raymond DiGiuseppe, Raymond Chip Tafrate, Understanding Anger Disorders, Oxford University Press, 2006, pp.133-159.

    [10] Anger,The American Heritage Dictionary of the English Language,Fourth Edition, 2000, Houghton Mifflin Company.

    [11] Michael Kent, Anger, The Oxford Dictionary of Sports Science &Medicine, Oxford University Press, ISBN 0-19-262845-3

    [12] Primate Ethology, 1967, Desmond Morris (Ed.). Weidenfeld & NicolsonPublishers: London, p.55

    [13] i Raymond W. Novaco, Anger, Encyclopedia of Psychology, OxfordUniversity Press, 2000

    [14] C.P.Browman and L.M.Goldstein. Towards an articulatory phonology.[15] W.N.Campbell and S.D.Isard.Segment durations in a syllable fra

    Suresh Kumar Thakur* et al. / (IJAEST) INTERNATIONAL JOURNAL OF ADVANCED ENGINEERING SCIENCES AND TECHNOLOGIESVol No. 8, Issue No. 2, 301 - 305

    ISSN: 2230-7818 @ 2011 http://www.ijaest.iserp.org. All rights Reserved. Page 305