Esta página está en construcción: perdonen los errores y temas inacabados.

This page is being developed: I am sorry for errors and unfinished subjects.

 

Automatic Recognition of Syntactic Intonation Patterns. (project)

 
3. WORK PROGRAMME
 
3.1 TITLE OF THE PROJECT:
 
Automatic Recognition of Syntactic Intonation Patterns.
 
3.2 OBJECTIVES (maximum half a page):
 
1. To recognize the intonation patterns in continuous speech. In principle, only syntactic speech, that is, speech devoid of expressive and impressive ‒pragmatical‒ intonation, is considered. One speaker ( at a time) is also considered for the beginning task.
 
2. To segment the acoustic chain in syntactical units, according to the recognition in 1. The main units considered are, from high to low: 1. Sentence.  2. Constituent (syntagma). 3. Syntagmatic word ‒one accented word with its clitc-like particles.
 
3. To punctuate the text, once it is obtained from the sound chain by means of phonetic and/or lexical speech recognition modules. Full stop will be attributed to sentences, commas to constituents; syntagmatic words will be not signed, except for internal analysis, but a graphic accent (‘) will be used.
 
 
3.3 BACKGROUND (maximum 2 pages):
 
Speech recognition is a difficult task due to ther variability in the actual realisation of all the abstract parameters and patterns in the language: phonemes, tonemes, expressions, syntactic structures, even meanings, appear in a myriad of forms which must be reduced to a small finite number of classes.
 
Among the parameters that play a role in the constitution of this complex structure, pitch ‒in the form of pitch patterns or melodies‒ share with timbre the predominant role. Indeed, meanwhile timbre or color ‒ linked to phonetics, phonemics, semantics‒ is responsible of a large part of  meanings (lexical), the main part of the rest (sentence meaning and sense) ‒ depends on syntax, which in turn is pure intonation; and the last part of the meaning ‒expressive, implications, sousentendus, irony‒ those strong modifiers of the general meaning are also mainly intonative, together with some other rhythmic characteristics.
 
Therefore, pitch estimation in a main task in any speech recognition system, either machinal, either human. And no system can be complete without it.
 
Automatic Pitch Estimation of a speech segment in not a very difficult task: periodicity measures, akin to autocorrelation techniques have been succesfully applied to it. But when the task consists on finding or recognizing patterns, few and robust intonative patterns, the job becomes harder.  First because the basic intonation patterns, the phonemical intonation patterns, have not been  not completely typified. Second due to the enormous variability that those patterns present. Third because often (almost always) those patterns present themselves intermingled ‒i.e. a doubtful assertion, an ironic acceptation, a rhethoric question‒.
 
It appears to be clear that while Intonation Recognition is a fundamental task, it is not an easy one. Our contribution in this project is described in the above and following paragraphs..
 
3.4 SCIENTIFIC AND TECHNICAL DESCRIPTION (approximately 3‑6 pages)
 
Measures of Speech Traits..
 
            Mathematically speaking, the period P (pitch) is the non‑zero value of the delay   J   for which the expression:
 
 
2. Pitch-Frequency.
 
                                             2 f (t+J) - f(t) 2
                        ADS(f) =    
                                          2 f (t+J) 2  -  2 f(t) 2    
 
                        is a minimum, below a threshold
 
                                                  2 f (t) 2  =  3  * fi *  
 
                         being the sum of the samples fi extended from  i=t+J/2  to  t+J/2.
 
 
3. Sonority‑Intensity.
                                   n
                      LI  = ( E  fi ) /  n
                                       i  
 
 
4. Quantity‑Duration.
 
                                           D = t1 ‑ t2,
 
                 t1= max ( d(LI) / dt),    t2 = max ( d(LI) / dt)
                          -t1                          -t2
 
Smoothing.
 
 
Distances.
 
A. Between Segments.
 
 
A. Between Polygonals.
 
 
3.4.1 RESEARCH ACTIVITIES:
 
Language (Semantics, Syntax, Phonetics, Prosody)
Metrics and Rhythm.
Microtonics
Non-western Musical Scales
Consonance Theory.
 
3.4.2 RESEARCH RESULTS:
 
Computer Aplications;
 
Bibliography.
 
Laboratory implementation.
 
3.5 MANAGEMENT INFORMATION
 
3.5.1 TASK DIVISION:
 
A. BUILDING THE MODEL
 
            1. Automatic pitch estimation of the speech segment.
            2. Smoothing and conexion of pitch segments.
            3. Simplification of pitch curve as a polygonal.
            4. Analysis of intonative patterns for the three syntactic units in the chosen language (English?).
            5. Time normalisation.
            6. Pairing patterns and syntactic units.
 
B. RECOGNITION.
 
            1. Automatic pitch estimation of the speech segment.
            2. Smoothing and conexion of pitch segments.
            3. Simplification of pitch curve as a polygonal.
            4. Time normalisation.
            5. Establishing:
 
                                     1. A distance measure between segments
                                     2. Adistance measure between patterns.
 
            6. Establishing a strategy to find the most likely pattern.
            7. Actual segmentation of the speech chain in syntactical units.
 
            8. Classification of each unit according to the model.
 
3.5.2 PLANNING:
 
3.5.3 EQUIPMENT:
 
3.5.3.1 Available.
3.5.3.2 Needed ( to adquire with project funds).
 
 
3.5.4 SCIENTIFIC REFERENCES:
 
3.5.4.1. PROSODY and INTONATION
 
Bergé, H. y Sánchez, F.J. (1995). Percepción tonal consciente e inconsciente en el habla. Castellón: Actas            Cong. AESLA95.
 
Sánchez, F.J.:"Application of Dissimilarity and Aperiodicity Functions to Pitch Extraction and Voiced‑ Unvoiced Decision. Proc. 9º ICA Cong., Madrid, 1977.
 
Sánchez, F.J.: Tratamiento Señales Pseudoperiódicas. Ph.D. Thesis Univ. Pol. Madrid, 1979.
 
Sánchez, F.J. and Bergé, H.: "Hacia una implementación computacional de una Semántica Multicontextual". Leng. naturales y leng. formales, IX. PPU. Barcelona, 1993.
 
Sánchez, F.J. (1995). Teoría preliminar de la sílaba  Lenguajes Naturales y Lenguajes Formales XI  C. Martín Vide (ed.)  Barcelona: Promociones y Publicaciones Universitarias: 569-578.
‒ (1996).  En busca de una interválica musical subyacente en la palabra. Lenguajes Naturales y Lenguajes Formales XII Martín Vide, C. (ed.)  Barcelona: Promociones y Publicaciones Universitarias: 539-546.
 
Sánchez, F.J. and  Caramés, J.L. (1996). Entaxis: a syntactical model of intonation. Abstr. II Int. Conf. Math. Linguistics. Tarragona: Univ. Rovira i Virgili: 91. Full text is going to appear soon, published by John Benjamins (1997).
 
Sánchez, F.J. y  Puyâ, M (1996).  Human and machinal recitation model for accentual (Spanish) and quantitative (Persian) metrics. Proc. ICEMCO 96 Cambridge: Cambridge University Press: 8.1.1-8.1.10
 
3.5.4.2. RECOGNITION
 
Fernández,R., Fúster,A., Jordá,S., Sánchez,F.J. "S.E.T.S. v.1B. Manual de Referencia". "S.E.T.S. v.1B. Manual  del  Usuario". "S.E.T.S. v.1. Listados  de  tratamiento  de  palabra". Lab.Tratamiento de Palabra y Música, I.E.C., C.S.I.C., Madrid, 1990.
 
 Sánchez, F.J, Fernández,R, Ponce,J, Trueba,C. "Memoria sobre reconocimiento de palabra". Informe anual del módulo 3 del  Proyecto "Cifrado...". Laboratorio de Tratamiento de Palabra y Música,  I.E.C., C.S.I.C.,  Madrid, 1990. Apén.1: Album Microfonético. Apen.2: Listados de las Aplicaciones.
 
Sánchez, F.J. "El motivo: apuntes de la asignatura Acústica III", Doctorado. Centro Superior para la Promoción de la Música, U.A.M., Madrid, 1990.
 
 Sánchez,F.J., Becerra, N. "Hacia la forma espectral relevante en el timbre de los alófonos sonoros del español (Castellano)", Simposio URSI, Cáceres, oct.1991. p.842.
 
 Sánchez, F.J., Becerra,N.J, Oñoro,A., Bergé,H.J. "Parametrización perceptiva en reconocimiento de palabra aislada", Simp.URSI, Málaga, sept.1992. p.269
 
Sánchez,F.J., Oñoro,A., Bergé,H.J. "Memoria sobre parametrización y reconocimiento de palabra. Informe anual del proyecto "Transcripción Texto a Habla", financiado por Comunidad de Madrid. Laboratorio de Tratamiento de Palabra y Música, I.E.C., C.S.I.C., Madrid, 1992.
 
 
4.   SUMMARY
 
4.1. TITLE OF THE PROJECT:
 
Automatic Recognition of Syntactic Intonation Patterns.
 
4.2. SUMMARY:
 
Speech recognition is a difficult task due to the variability in the actual realisation of all the abstract parameters and patterns in the language: phonemes, tonemes, expressions, syntactic structures, even meanings. Pitch ‒in the form of pitch patterns or melodies‒ has a  predominant role: sentence meaning and sense depends on syntax, which in turn depends on intonation, as well as pragmatic expressions. Therefore, pitch estimation is a main task in any speech recognition system, either machinal, either human. Automatic Pitch Estimation of Intonation Patterns will be realized by means of periodicity measures, akin to autocorrelation techniques. After, the task consists in recognizing patterns, few and robust intonative patterns. The system will therefore accomplish the following subtasks: 1. To recognize the intonation patterns in continuous speech. 2. To segment the acoustic chain in syntactical units (a. Sentence. b. Constituent (syntagma). c. Syntagmatic word ‒one accented word with its clitic-like particles). 3. To punctuate the text, once it is obtained by the  phonetic/lexical  recognition modules.
 
 

Vuelta al Principio     Última actualización :  viernes, 18 de septiembre de 2015 Visitantes: contador de visitas