Publication: Reconocimiento de habla mediante transparametrización : una alternativa robusta para entornos móviles e IP
Loading...
Identifiers
Publication date
2002
Defense date
2002-02-05
Authors
Advisors
Tutors
Journal Title
Journal ISSN
Volume Title
Publisher
Impact
Export
Abstract
En el panorama actual de las telecomunicaciones, dos son los tipos de redes con mayor éxito en la actualidad: las redes de móviles y las redes de paquetes basadas en el protocolo TCP/IP (–Transport Control Protocol / Internet Protocol-). Entre los factores que han llevado al éxito de las primeras en su segunda generación (2G) está su ubicuidad, es decir, gracias al enorme despliegue geográfico de estas redes es posible
realizar una llamada telefónica desde casi cualquier localización (en el mundo desarrollado). Por su parte, las redes IP (originalmente diseñadas para el transporte de
datos) también están logrando imponer su presencia en detrimento de cualquier otro tipo
de red fija y uno de sus puntos fuertes es, sin duda, su capacidad –todavía bastante limitada– para transmitir cualquier tipo de información multimedia.
Uno de los puntos de convergencia entre las dos redes es su objetivo de permitir que todo tipo de información transite por ellas con ciertas garantías de calidad de servicio (QoS -Quality of Service-). Esto está motivado por la cantidad de nuevas aplicaciones que pueden crearse a partir de la posibilidad de combinar informaciones de distinto tipo (texto, video, voz, imágenes, música, etc.) y las tecnologías del habla están llamadas a jugar un papel fundamental a través del desarrollo de interfaces más naturales para estas aplicaciones.
Entre estas tecnologías, el reconocimiento de habla está llegando a una fase de madurez que hace cada vez más viables estos desarrollos. De hecho, desde hace algún tiempo se viene prestando mucha atención a la robustez de estos sistemas cuando se trasladan al mundo real, habiéndose desarrollado numerosas técnicas para enfrentarse a problemas tales como: variaciones en el entorno acústico, influencia de los
transductores y el canal de transmisión y variaciones en el hablante y la tarea que se
aborda.
En esta tesis estudiamos la influencia de dos tipos de canales de transmisión concretos, representantes de los dos tipos de redes que hemos venido introduciendo: el estándar europeo para comunicaciones móviles GSM (-Global System for Mobile-, anteriormente –Group Speciale Mobile) y el de las actuales redes basadas en los
protocolos TCP/IP. Además, proponemos una solución, que hemos denominado reconocimiento mediante transparametrización, con la que mejoramos la tasas de reconocimiento en ambos entornos y que, aunque en un principio, hemos
particularizado para dichos entornos, puede ser aplicada en otros.
La característica común de la transmisión de voz a través de estas dos redes es el proceso de codificación que tiene lugar para adecuar su régimen binario reducir. Esta compresión con pérdidas de la señal de voz produce un deterioro de su calidad, que si bien es aceptable en el caso de reconocedores humanos –los codificadores están diseñados para minimizar la distorsión perceptible–, se traduce en una disminución
apreciable de las prestaciones de los reconocedores automáticos.
Por otra parte, los errores de transmisión que se producen en ambos entornos, contribuyen también a la degradación de las prestaciones de los reconocedores. En GSM, estos errores aparecen en forma de ráfagas de bits erróneos producidas por desvanecimientos de la señal de radiofrecuencia, que pueden afectar a una o varias tramas consecutivas, completamente o sólo en parte. El caso de IP es algo distinto ya
que, en general, no se suelen producir errores de bit a ráfagas y muy raramente errores
aislados, debido a la alta fiabilidad del canal, sino que lo más común es que se produzcan pérdidas de paquetes (a ráfagas) en los nodos de enrutamiento.
En cualquier caso, lo que se pone de manifiesto en esta tesis es que el hecho de que este tipo de errores se produzcan sobre la voz codificada tiene consecuencias que no se pueden tratar de la misma manera que si se produjeran sobre la señal de voz original (por ejemplo, modelando los errores haciendo las hipótesis habituales de ruido convolutivo o aditivo). Es decir, si tenemos en cuenta que el proceso de codificación de la voz consiste, a grandes rasgos, en la extracción de una serie de parámetros que
representan distintos aspectos específicos de este tipo de señal (su periodo fundamental,
la posición de sus formantes, su característica sonora o sorda, su energía, etc), nos
percataremos de que la modificación de cada uno de ellos tiene consecuencias muy distintas sobre la señal vocal reconstruida.
Un reconocedor convencional que recibe una señal de voz codificada, la primera acción que realiza sobre ella es su decodificación y de esa forma, ya puede proceder a realizar la extracción de características o parametrización para reconocimiento. En este proceso, las distorsiones de codificación y de los errores se trasladan a los parámetros a partir de los que se realizará el reconocimiento, produciendo el deterioro de las prestaciones del reconocedor.
Para mejorar esta situación, en esta tesis proponemos el análisis de la parametrización de la señal de voz que lleva a cabo el codificador antes de su decodificación y la transformación de ésta en otra adecuada para el reconocimiento.
Esto además, nos permite utilizar métodos de recuperación frente a errores y de transformación de parametrizaciones orientados directamente al reconocimiento, sin limitarnos a los ya previstos en los estándares de codificación, cuyo propósito es recuperar una señal de voz perceptualmente aceptable sujetos a una fuerte restricción de tiempo real.
De esta forma, obtenemos una solución aplicable a ambos entornos (GSM e IP) que reduce la influencia sobre los reconocedores, por una parte, la distorsión de codificación haciendo una selección de la información relevante para reconocimiento, y por otra, el efecto de los errores de transmisión, actuando directamente sobre los parámetros afectados. Resulta notable el hecho de que esta solución sea aplicable tanto a entornos móviles como a redes de tipo IP, ya que puede emplearse cuando existe una
combinación de ambas como parece ser la tendencia.
_________________________________________________
Nowadays, two are the most important types of telecommunication networks worldwide: mobile networks and those based on the TCP/IP (–Transport Control Protocol / Internet Protocol-) protocol. Among the factors which have led the first ones to success in its second generation (2G) is its ubiquity, i.e., the enormous geographic deployment of these networks makes it possible to place a phone call virtually from any location (in the developed world). By its side, IP networks (originally designed for data transport) are succeeding in progressively substituting other kinds of fixed networks and one of its most remarkable advantages is, no doubt, its ability –still quite limited– for transmitting any kind of multimedia information. One of the points of convergence between both networks is thus, its aim to provide the means for any kind of information to be sent over it with a certain quality of service (QoS). This is motivated by the huge range of new applications that can be created due to the possibility of combining multiple kind of information types (text, video, voice, images, music, etc.), and here is where speech technologies are called to play a fundamental role by providing more natural interfaces. Among these technologies, speech recognition is attaining enough maturity to allow for these developments. In fact, during the last years, much attention has been payed to the robustness of these systems when they are transferred to the real world, having many techniques been developed to cope with problems such as: acoustic environment variations, transductors and transmission channels influences and variations in the speaker and tasks. In this thesis we will analyse the influence of two specific transmission channel types, which represent an example of both types of network introduced above: European mobile communications standard, GSM (Global System for Mobile, which previously stood for –Group Speciale Mobile-) and the present-day TCP/IP-based networks. Besides, we propose an alternative solution which we have named by which we attain improvements in the performances of recognizers under both environments conditions and, though tested specifically in those particular networks, also potencially applicable in some other. A common characteristic of speech transmission over these networks is the previous speech coding process that takes place to adequate the bit rate. This lossy process produces a quality drop which, though not very disappointing for human recognizers –as codecs are designed to minimize perceptual distortion–, sometimes harmful for automatic speech recognizers. Besides, transmission errors caused by both networks, contribute to the degradation of the recognizers accuracy. Under the GSM environment, these errors appear in the form of bit bursts produced by signal fadings in radio-frequency channels which can affect one or more consecutive frames, just in part or completely. IP-based networks is somehow different because, generally speaking, isolated or burst bit errors are not usually encountered due to the high channel reliability, but on the contrary, it is common for the router nodes to drop packets during congestion situations. In any case, what is here emphasized is that the fact that when these errors are placed on the coded speech signal produce effects that cannot be treated the same way as if they had been produced on the original one (for example, modelling error sources as convolutive or additive). In other words, if we take into account that the speech coding process consists, roughly speaking, on the extraction of a series of parameters representing the various aspects of this particular kind of signal (its fundamental frequency, formant positions, voicing mode, energy, etc.), it will become apparent that the modification of each of them affects in a different manner the reconstructed speech. A conventional speech recognizer receiving a coded signal, first proceeds to its decoding which enables the usual feature extraction procedure that recognizers habitually perform. By this process, coding and errors distortions are transferred to the recognition parameterization, finally producing a disminishing of the recognition rates. To improve this situation, we propose the analysis of the speech codec parameterization before its decoding and its transformation into an adequate one for recognition. This, besides, allows us to use recognition-oriented error recovery and parameterization transformation methods not restricted to the ones provided by the coding standards, whose aim is to recover perceptually acceptable signals subject to very strict real-time restrictions. Therefore, we have obtained a method applicable to both environments, by which we are capable of reducing the influence on the recognizer by minimizing, on the one hand, the coding distortion by means of making a selection of the relevant recognition information embedded into the speech coding parameterization and, on the other hand, the effect of the transmission errors, by acting straight on the affected parameters. It is worth mention that the fact that this method is effective both under mobile and IP networks environments can represent an advantage for its application in combined situations, which seems to be the trend.
Nowadays, two are the most important types of telecommunication networks worldwide: mobile networks and those based on the TCP/IP (–Transport Control Protocol / Internet Protocol-) protocol. Among the factors which have led the first ones to success in its second generation (2G) is its ubiquity, i.e., the enormous geographic deployment of these networks makes it possible to place a phone call virtually from any location (in the developed world). By its side, IP networks (originally designed for data transport) are succeeding in progressively substituting other kinds of fixed networks and one of its most remarkable advantages is, no doubt, its ability –still quite limited– for transmitting any kind of multimedia information. One of the points of convergence between both networks is thus, its aim to provide the means for any kind of information to be sent over it with a certain quality of service (QoS). This is motivated by the huge range of new applications that can be created due to the possibility of combining multiple kind of information types (text, video, voice, images, music, etc.), and here is where speech technologies are called to play a fundamental role by providing more natural interfaces. Among these technologies, speech recognition is attaining enough maturity to allow for these developments. In fact, during the last years, much attention has been payed to the robustness of these systems when they are transferred to the real world, having many techniques been developed to cope with problems such as: acoustic environment variations, transductors and transmission channels influences and variations in the speaker and tasks. In this thesis we will analyse the influence of two specific transmission channel types, which represent an example of both types of network introduced above: European mobile communications standard, GSM (Global System for Mobile, which previously stood for –Group Speciale Mobile-) and the present-day TCP/IP-based networks. Besides, we propose an alternative solution which we have named by which we attain improvements in the performances of recognizers under both environments conditions and, though tested specifically in those particular networks, also potencially applicable in some other. A common characteristic of speech transmission over these networks is the previous speech coding process that takes place to adequate the bit rate. This lossy process produces a quality drop which, though not very disappointing for human recognizers –as codecs are designed to minimize perceptual distortion–, sometimes harmful for automatic speech recognizers. Besides, transmission errors caused by both networks, contribute to the degradation of the recognizers accuracy. Under the GSM environment, these errors appear in the form of bit bursts produced by signal fadings in radio-frequency channels which can affect one or more consecutive frames, just in part or completely. IP-based networks is somehow different because, generally speaking, isolated or burst bit errors are not usually encountered due to the high channel reliability, but on the contrary, it is common for the router nodes to drop packets during congestion situations. In any case, what is here emphasized is that the fact that when these errors are placed on the coded speech signal produce effects that cannot be treated the same way as if they had been produced on the original one (for example, modelling error sources as convolutive or additive). In other words, if we take into account that the speech coding process consists, roughly speaking, on the extraction of a series of parameters representing the various aspects of this particular kind of signal (its fundamental frequency, formant positions, voicing mode, energy, etc.), it will become apparent that the modification of each of them affects in a different manner the reconstructed speech. A conventional speech recognizer receiving a coded signal, first proceeds to its decoding which enables the usual feature extraction procedure that recognizers habitually perform. By this process, coding and errors distortions are transferred to the recognition parameterization, finally producing a disminishing of the recognition rates. To improve this situation, we propose the analysis of the speech codec parameterization before its decoding and its transformation into an adequate one for recognition. This, besides, allows us to use recognition-oriented error recovery and parameterization transformation methods not restricted to the ones provided by the coding standards, whose aim is to recover perceptually acceptable signals subject to very strict real-time restrictions. Therefore, we have obtained a method applicable to both environments, by which we are capable of reducing the influence on the recognizer by minimizing, on the one hand, the coding distortion by means of making a selection of the relevant recognition information embedded into the speech coding parameterization and, on the other hand, the effect of the transmission errors, by acting straight on the affected parameters. It is worth mention that the fact that this method is effective both under mobile and IP networks environments can represent an advantage for its application in combined situations, which seems to be the trend.
Description
Keywords
Reconocimiento de voz, Redes de comunicación, Codificación de voz