Inigem

The present work consists on the use of delta cepstra coefficients in Mel scale, wavelet and wavelet packet transforms to feed a system for automatic speaker identification based on neural networks. Different alternatives are tested for the classifier based on neural nets, having achieved very good performance for closed groups of speakers in a text independent form. When a single neural net is used for all the speakers, the results decay abruptly with increasing number of speakers to identify. This takes to implement a system where there is one neural net for each speaker, which provided excellent results, compared with the opposing ones in the bibliography using other methods. This classifier structure possesses other advantages, for example, adding a new speaker to the system only requires to train a net for the speaker in question, in contrast with a system where the classifier is formed by a single great net, which should be in general be trained completely again.

The detection of changes in the parameter values of a nonlinear dynamic system is a branch of study with multiple applications. In this paper, we explore a variant of an automatic detector and clustering of slight parameter variations in nonlinear dynamic systems proposed by Torres et al. [Automatic detection of slight changes in nonlinear dynamical systems using multiresolution entropy tools, Int. J. Bifurc. Chaos 11(4) (2001) 967–981]. The new method takes the advantages of the continuous multiresolution entropy to localize slight changes in the parameters, and uses self-organizing maps to quantify and cluster these changes. We discuss the performance of this method while applied to automatic segmentation of natural and synthetic diphthongs in the presence of additive noise. Our results show the potentiality of the proposed method.

This project involved the design and development of a relational SQL-based database to generate an intonational model for an Argentine Spanish text to speech system. The first stage in the population of the database involved the massive loading of text, divided into three co-indexed files: sentences, orthographic words and phonological syllables. A software tool, which performed phonemic transcription and syllabic segmentation of the text, was developed to allow indexation. In the beginning, a large set of sentences was loaded, then, a subset of 741 sentences was selected, according to criteria related to syllable occurrences in all positions in words with and without stress. This set contained 97% of all Spanish syllables extracted from a widely used Spanish dictionary. The utterances were recorded at 16 kHz / 16 bits, using an interactive program. Two professional announcers were instructed to generate a variety of accent patterns and intonational phrases to prevent monotony. Speech signals were then labeled by trained phoneticians with a spectral analysis tool, using ToBI tiers (Beckman & Ayers, 1994). ToBI conventions were adopted to account for the prosodic patterns of Argentine Spanish. Frequency values were scaled using the ERB scale, and bitonal accents were redefined. Finally, the second stage of the population of the database consisted of the incorporation of the labeled files. Waveforms were kept outside the database, but linked to it, to allow the identification, reproduction, and selection of specific segments.

This work evaluates the efficiency of different word classes -part of speech-, normalized vs. non normalized counting for syllable and word occurrences, to predict non orthographic breaks of an Argentine Spanish database, designed for the development of the prosody component for a Text To Speech system. Within a set of 741 sentences, regression trees were trained and tested with two different proportions of data. The results show an error range of 8 to 15% whose minimum value is related to a reduced amount of morphologic categories, and a normalized counting of syllables  and words.

We evaluate here the application of two intonational models –quantitative and phonetic- to the analysis of an Argentine Spanish database of 741 broad-focus declarative sentences. The analytic model is the superpositional model proposed by Fujisaki (2003) for several languages. The phonetic model is the result of the application of a labelling method (Gurlekian et al., 2001b) that incorporates psycho-acoustic measurements and a detailed description of the shape of the accent. Parameters generated by this labelling method were used to synthesize the intonational contours, which were then evaluated in a perception test. Results indicate the validity of Fujisaki´s model for describing a large database of Spanish broad-focus declaratives, and, thus, suggest the importance of Fujisaki’s model for speech technology applications. The extended ToBI model (ToBI-A) is validated by the correlation coefficient and RMSE values as well as the results of the perception test. Ten native speakers of the variety under study judged the synthesized sentences as highly natural with only minor differences with the original contour. These results indicate that the ToBI-A (i) is adequate for a linguistically-meaningful description of the intonation of a new variety; (ii) can be adequately used for modelling intonation.

The goal of this project was the design and realisation of a database to be used in an automatic speech recognition system for a fixed telephone network. One thousand speakers, native to five Argentine dialectal regions, were recorded. Each speaker answered five questions and read 38 texts, which consisted of numbers, names, last names, corporation names, and phonetically rich sentences. Two hundred sets of nine sentences, selected from a pool of 7000 sentences, were used to generate 1,000 prompt sheets. These sentences contained all of the contextual allophones transcribed with the SAMPA alphabet. A strategy of uniform information collection and recording was designed. Phone calls were received through a digital line connected to a computer, and monitored by acquisition software. This procedure resulted in an efficient way to collect data from a large number of speakers representing several different dialects in only two months.

Synthesis by concatenation of natural speech improves perceptual results when phonemes and syllables are segmented at places where spectral variations are small [Klatt, D., 1987. Review of text-to-speech conversion for English. J. Acoust. Soc. Am 82 (3), 737–793]. An automatic segmentation method is explored here, using a tool based on a combination of Entropy Coding, Multiresolution Analysis, and Kohonen’s Self Organized Maps. The segmentation method considers that there are no limits imposed by any linguistic unit. Resulting waveforms represent phone chains dominated by spectral dynamic structures. Each acoustic unit obtained could be composed of a variety of phonemes or a segmented part of them at the unit boundary. The number of units and unit structure are speaker dependent, i.e. rate, segmental and suprasegmental distinctive features affect them as dynamic structure varies. Results obtained from two databases – one male, one female – of 741 sentences each show this dependence, presenting a different number of units and occurrences for each speaker. Nevertheless, both speakers show a high occurrence of three (36–24%) and four (29–27%) phoneme sequences. Vowel-consonant- vowel sequences are the most frequent type (9.7–8.3%). Consonant-vowel syllables, which are phonemically frequent in Spanish (58%), are less represented (6.6–3.2%) using this method. The relevance of half phone segmentation is verified given that 66% for the female speaker and 53% for the male speaker, of the total units start and end with a segmented phone. Perceptual experiments showed that concatenated speech, created with dynamic acoustic units, was judged more natural than with diphone units.

Los paquetes de onditas (WP) son una extensión de la transformada de onditas (WT) introducida recientemente. Si bien su aplicación ha sido amplia en problemas de compresión, filtrado y supresión de ruido, poco se ha hecho en el problema de clasificación de señales. Los fonemas del habla constituyen un ejemplo de señales de duración variable de dificil separación. En este trabajo se presentan los resultados de la aplicación de paquetes de onditas diseñados en base a criterios perceptuales para el análisis y la clasificación de fonemas. Debido a la naturaleza dinámica de los patrones el clasificador utilizado es una red neuronal con retardos temporales (TDNN).

El objetivo de este trabajo es el diseño y la aplicación de una prueba rápida de inteligibilidad del habla en ambientes ruidosos. La prueba está diseñada para niños de 6 a 12 años que asisten a escuelas o instituciones donde no se dispone de equipos de audiometría e instalaciones preparadas para controlar el ruido-ambiente. Se evaluará la capacidad de identificar palabras —mediante auriculares— pronunciadas en el contexto de oraciones en condiciones controladas respecto del ruido. La prueba se presenta al nivel que requiere cada niño para obtener un reconocimiento del 100% de las palabras en silencio. Luego, las oraciones y el ruido enmascarador se presentan de forma simultánea en el mismo oído a relaciones señal a ruido que van desde 20 a -5 dB. El ruido es el denominado ruido modulado con forma de habla (Modulated Speech Shaped Noise, MSSN), que ha demostrado ser muy efectivo como enmascarador. Los resultados para niños con audición normal indican 1) el nivel de presión sonora para el reconocimiento del 100% de las palabras en silencio, medido en el oído artificial es en promedio de 33,9 dB (desviación: 0,9) en el laboratorio y de 36,77 dB (desviación:1,8) en la escuela, y 2) el 50% de identificación de la palabra se alcanza a la relación señal a ruido de 0 dB tanto en el laboratorio como en la escuela. Se puede concluir que la prueba permitirá evaluaciones rápidas de la inteligibilidad en lugares con diferente nivel de ruido-ambiente y sin disponibilidad de recursos técnicos.

This work presents two new approaches for parameter estimation of the superpositional intonation model for German. These methods introduce linguistic assumptions, which allow to initialize a previous standard method. Also, eliminates restrictions on the configuration of accents. The proposed linguistic hypotheses can be used based on tonal or lexical accent, which gives rise to two different estimation methods. These hypotheses were validated by comparison of the estimation performance relative to two standard methods, one manual and one automatic. The results show that the proposed methods far exceed the performance of the automatic and slightly beyond the manual method of reference.

In this paper we model the segmental duration of Spanish spoken in Buenos Aires, considering its application in a text-to-speech system. The work was performed on two hand labeled databases. We use artificial neural networks as predictor, and all the input features can be extracted automatically from the speech text. We experimented with a neural network for all phonemes and one neural network for phoneme. In both cases the results are very promising for the two databases used. The order of importance of input features revealed to be different for each of the methods tested and different according to the speaker style.

This paper introduces Aromo text-to-speech system for Argentine Spanish, which was designed for telephony applications and is based on unit selection and concatenation. The system operates as a client-server engine that supports MRCP, SIP and SSML technologies. Perceptual evaluation results show that Aromo’s voice achieve high performances in both naturalness and intelligibility.

El presente trabajo consiste en la utilización de Coeficientes Delta Cepstra en escala de Mel para alimentar un sistema de identificación automática del hablante basado en redes neuronales. Se ensayan diferentes alternativas para el clasificador basado en redes neuronales, lográndose muy buenos desempeños para conjuntos cerrados de hablantes en forma independiente del texto.