Phonotactic language recognition using i-vectors and phoneme posteriogram counts

This paper describes a novel approach to phonotactic LID, where instead of using soft-counts based on phoneme lattices, we use posteriogram to obtain n-gram counts. The high-dimensional vectors of counts are reduced to low-dimensional units for which we adapted the commonly used term i-vectors. The reduction is based on multinomial subspace modeling and is designed to work in the total-variability space. The proposed technique was tested on the NIST 2009 LRE set with better results to a system based on using soft-counts (Cavg on 30s: 3.15% vs 3.43%), and with very good results when fused with an acoustic i-vector LID system (Cavg on 30s acoustic 2.4% vs 1.25%). The proposed technique is also compared with another low dimensional projection system based on PCA. In comparison with the original soft-counts, the proposed technique provides better results, reduces the problems due to sparse counts, and avoids the process of using pruning techniques when creating the lattices.

Evaluación de la contribución y el impacto de las tecnologías del habla en la detección automática del Síndrome de la Apnea Obstructiva del Sueño = Contributions and impact assessment of speech technologies on the automatic detection of severe Obstructive Sleep Apnea syndrome

La presente Tesis analiza las posibilidades que ofrecen en la actualidad las tecnologías del habla para la detección de patologías clínicas asociadas a la vía aérea superior. El estudio del habla que tradicionalmente cubre tanto la producción como el proceso de transformación del mensaje y las señales involucradas, desde el emisor hasta alcanzar al receptor, ofrece una vía de estudio alternativa para estas patologías. El hecho de que la señal emitida no solo contiene este mensaje, sino también información acerca del locutor, ha motivado el desarrollo de sistemas orientados a la identificación y verificación de la identidad de los locutores. Estos trabajos han recibido recientemente un nuevo impulso, orientándose tanto hacia la caracterización de rasgos que son comunes a varios locutores, como a las diferencias existentes entre grabaciones de un mismo locutor. Los primeros resultan especialmente relevantes para esta Tesis dado que estos rasgos podrían evidenciar la presencia de características relacionadas con una cierta condición común a varios locutores, independiente de su identidad. Tal es el caso que se enfrenta en esta Tesis, donde los rasgos identificados se relacionarían con una de la patología particular y directamente vinculada con el sistema de físico de conformación del habla. El caso del Síndrome de Apneas Hipopneas durante el Sueno (SAHS) resulta paradigmático. Se trata de una patología con una elevada prevalencia mundo, que aumenta con la edad. Los pacientes de esta patología experimentan episodios de cese involuntario de la respiración durante el sueño, que se prolongan durante varios segundos y que se reproducen a lo largo de la noche impidiendo el correcto descanso. En el caso de la apnea obstructiva, estos episodios se deben a la imposibilidad de mantener un camino abierto a través de la vía aérea, de forma que el flujo de aire se ve interrumpido. En la actualidad, el diagnostico de estos pacientes se realiza a través de un estudio polisomnográfico, que se centra en el análisis de los episodios de apnea durante el sueño, requiriendo que el paciente permanezca en el hospital durante una noche. La complejidad y el elevado coste de estos procedimientos, unidos a las crecientes listas de espera, han evidenciado la necesidad de contar con técnicas rápidas de detección, que si bien podrían no obtener tasas tan elevadas, permitirían reorganizar las listas de espera en función del grado de severidad de la patología en cada paciente. Entre otros, los sistemas de diagnostico por imagen, así como la caracterización antropométrica de los pacientes, han evidenciado la existencia de patrones anatómicos que tendrían influencia directa sobre el habla. Los trabajos dedicados al estudio del SAHS en lo relativo a como esta afecta al habla han sido escasos y algunos de ellos incluso contradictorios. Sin embargo, desde finales de la década de 1980 se conoce la existencia de patrones específicos relativos a la articulación, la fonación y la resonancia. Sin embargo, su descripción resultaba difícilmente aprovechable a través de un sistema de reconocimiento automático, pero apuntaba la existencia de un nexo entre voz y SAHS. En los últimos anos las técnicas de procesado automático han permitido el desarrollo de sistemas automáticos que ya son capaces de identificar diferencias significativas en el habla de los pacientes del SAHS, y que los distinguen de los locutores sanos. Por contra, poco se conoce acerca de la conexión entre estos nuevos resultados, los sé que habían obtenido en el pasado y la patogénesis del SAHS. Esta Tesis continua la labor desarrollada en este ámbito considerando específicamente: el estudio de la forma en que el SAHS afecta el habla de los pacientes, la mejora en las tasas de clasificación automática y la combinación de la información obtenida con los predictores utilizados por los especialistas clínicos en sus evaluaciones preliminares. Las dos primeras tareas plantean problemas simbióticos, pero diferentes. Mientras el estudio de la conexión entre el SAHS y el habla requiere de modelos acotados que puedan ser interpretados con facilidad, los sistemas de reconocimiento se sirven de un elevado número de dimensiones para la caracterización y posterior identificación de patrones. Así, la primera tarea debe permitirnos avanzar en la segunda, al igual que la incorporación de los predictores utilizados por los especialistas clínicos. La Tesis aborda el estudio tanto del habla continua como del habla sostenida, con el fin de aprovechar las sinergias y diferencias existentes entre ambas. En el análisis del habla continua se tomo como punto de partida un esquema que ya fue evaluado con anterioridad, y sobre el cual se ha tratado la evaluación y optimización de la representación del habla, así como la caracterización de los patrones específicos asociados al SAHS. Ello ha evidenciado la conexión entre el SAHS y los elementos fundamentales de la señal de voz: los formantes. Los resultados obtenidos demuestran que el éxito de estos sistemas se debe, fundamentalmente, a la capacidad de estas representaciones para describir dichas componentes, obviando las dimensiones ruidosas o con poca capacidad discriminativa. El esquema resultante ofrece una tasa de error por debajo del 18%, sirviéndose de clasificadores notablemente menos complejos que los descritos en el estado del arte y de una única grabación de voz de corta duración. En relación a la conexión entre el SAHS y los patrones observados, fue necesario considerar las diferencias inter- e intra-grupo, centrándonos en la articulación característica del locutor, sustituyendo los complejos modelos de clasificación por el estudio de los promedios espectrales. El resultado apunta con claridad hacia ciertas regiones del eje de frecuencias, sugiriendo la existencia de un estrechamiento sistemático en la sección del tracto en la región de la orofaringe, ya prevista en la patogénesis de este síndrome. En cuanto al habla sostenida, se han reproducido los estudios realizados sobre el habla continua en grabaciones de la vocal /a/ sostenida. Los resultados son cualitativamente análogos a los anteriores, si bien en este caso las tasas de clasificación resultan ser más bajas. Con el objetivo de identificar el sentido de este resultado se reprodujo el estudio de los promedios espectrales y de la variabilidad inter e intra-grupo. Ambos estudios mostraron importantes diferencias con los anteriores que podrían explicar estos resultados. Sin embargo, el habla sostenida ofrece otras oportunidades al establecer un entorno controlado para el estudio de la fonación, que también había sido identificada como una fuente de información para la detección del SAHS. De su estudio se pudo observar que, en el conjunto de datos disponibles, no existen variaciones que pudieran asociarse fácilmente con la fonación. Únicamente aquellas dimensiones que describen la distribución de energía a lo largo del eje de frecuencia evidenciaron diferencias significativas, apuntando, una vez más, en la dirección de las resonancias espectrales. Analizados los resultados anteriores, la Tesis afronta la fusión de ambas fuentes de información en un único sistema de clasificación. Con ello es posible mejorar las tasas de clasificación, bajo la hipótesis de que la información presente en el habla continua y el habla sostenida es fundamentalmente distinta. Esta tarea se realizo a través de un sencillo esquema de fusión que obtuvo un 88.6% de aciertos en clasificación (tasa de error del 11.4%), lo que representa una mejora significativa respecto al estado del arte. Finalmente, la combinación de este clasificador con los predictores utilizados por los especialistas clínicos ofreció una tasa del 91.3% (tasa de error de 8.7%), que se encuentra dentro del margen ofrecido por esquemas más costosos e intrusivos, y que a diferencia del propuesto, no pueden ser utilizados en la evaluación previa de los pacientes. Con todo, la Tesis ofrece una visión clara sobre la relación entre el SAHS y el habla, evidenciando el grado de madurez alcanzado por la tecnología del habla en la caracterización y detección del SAHS, poniendo de manifiesto que su uso para la evaluación de los pacientes ya sería posible, y dejando la puerta abierta a futuras investigaciones que continúen el trabajo aquí iniciado. ABSTRACT This Thesis explores the potential of speech technologies for the detection of clinical disorders connected to the upper airway. The study of speech traditionally covers both the production process and post processing of the signals involved, from the speaker up to the listener, offering an alternative path to study these pathologies. The fact that utterances embed not just the encoded message but also information about the speaker, has motivated the development of automatic systems oriented to the identification and verificaton the speaker’s identity. These have recently been boosted and reoriented either towards the characterization of traits that are common to several speakers, or to the differences between records of the same speaker collected under different conditions. The first are particularly relevant to this Thesis as these patterns could reveal the presence of features that are related to a common condition shared among different speakers, regardless of their identity. Such is the case faced in this Thesis, where the traits identified would relate to a particular pathology, directly connected to the speech production system. The Obstructive Sleep Apnea syndrome (OSA) is a paradigmatic case for analysis. It is a disorder with high prevalence among adults and affecting a larger number of them as they grow older. Patients suffering from this disorder experience episodes of involuntary cessation of breath during sleep that may last a few seconds and reproduce throughout the night, preventing proper rest. In the case of obstructive apnea, these episodes are related to the collapse of the pharynx, which interrupts the air flow. Currently, OSA diagnosis is done through a polysomnographic study, which focuses on the analysis of apnea episodes during sleep, requiring the patient to stay at the hospital for the whole night. The complexity and high cost of the procedures involved, combined with the waiting lists, have evidenced the need for screening techniques, which perhaps would not achieve outstanding performance rates but would allow clinicians to reorganize these lists ranking patients according to the severity of their condition. Among others, imaging diagnosis and anthropometric characterization of patients have evidenced the existence of anatomical patterns related to OSA that have direct influence on speech. Contributions devoted to the study of how this disorder affects scpeech are scarce and somehow contradictory. However, since the late 1980s the existence of specific patterns related to articulation, phonation and resonance is known. By that time these descriptions were virtually useless when coming to the development of an automatic system, but pointed out the existence of a link between speech and OSA. In recent years automatic processing techniques have evolved and are now able to identify significant differences in the speech of OSAS patients when compared to records from healthy subjects. Nevertheless, little is known about the connection between these new results with those published in the past and the pathogenesis of the OSA syndrome. This Thesis is aimed to progress beyond the previous research done in this area by addressing: the study of how OSA affects patients’ speech, the enhancement of automatic OSA classification based on speech analysis, and its integration with the information embedded in the predictors generally used by clinicians in preliminary patients’ examination. The first two tasks, though may appear symbiotic at first, are quite different. While studying the connection between speech and OSA requires simple narrow models that can be easily interpreted, classification requires larger models including a large number dimensions for the characterization and posterior identification of the observed patterns. Anyhow, it is clear that any progress made in the first task should allow us to improve our performance on the second one, and that the incorporation of the predictors used by clinicians shall contribute in this same direction. The Thesis considers both continuous and sustained speech analysis, to exploit the synergies and differences between them. On continuous speech analysis, a conventional speech processing scheme, designed and evaluated before this Thesis, was taken as a baseline. Over this initial system several alternative representations of the speech information were proposed, optimized and tested to select those more suitable for the characterization of OSA-specific patterns. Evidences were found on the existence of a connection between OSA and the fundamental constituents of the speech: the formants. Experimental results proved that the success of the proposed solution is well explained by the ability of speech representations to describe these specific OSA-related components, ignoring the noisy ones as well those presenting low discrimination capabilities. The resulting scheme obtained a 18% error rate, on a classification scheme significantly less complex than those described in the literature and operating on a single speech record. Regarding the connection between OSA and the observed patterns, it was necessary to consider inter-and intra-group differences for this analysis, and to focus on the articulation, replacing the complex classification models by the long-term average spectra. Results clearly point to certain regions on the frequency axis, suggesting the existence of a systematic narrowing in the vocal tract section at the oropharynx. This was already described in the pathogenesis of this syndrome. Regarding sustained speech, similar experiments as those conducted on continuous speech were reproduced on sustained phonations of vowel / a /. Results were qualitatively similar to the previous ones, though in this case perfomance rates were found to be noticeably lower. Trying to derive further knowledge from this result, experiments on the long-term average spectra and intraand inter-group variability ratios were also reproduced on sustained speech records. Results on both experiments showed significant differences from the previous ones obtained from continuous speech which could explain the differences observed on peformance. However, sustained speech also provided the opportunity to study phonation within the controlled framework it provides. This was also identified in the literature as a source of information for the detection of OSA. In this study it was found that, for the available dataset, no sistematic differences related to phonation could be found between the two groups of speakers. Only those dimensions which relate energy distribution along the frequency axis provided significant differences, pointing once again towards the direction of resonant components. Once classification schemes on both continuous and sustained speech were developed, the Thesis addressed their combination into a single classification system. Under the assumption that the information in continuous and sustained speech is fundamentally different, it should be possible to successfully merge the two of them. This was tested through a simple fusion scheme which obtained a 88.6% correct classification (11.4% error rate), which represents a significant improvement over the state of the art. Finally, the combination of this classifier with the variables used by clinicians obtained a 91.3% accuracy (8.7% error rate). This is within the range of alternative, but costly and intrusive schemes, which unlike the one proposed can not be used in the preliminary assessment of patients’ condition. In the end, this Thesis has shed new light on the underlying connection between OSA and speech, and evidenced the degree of maturity reached by speech technology on OSA characterization and detection, leaving the door open for future research which shall continue in the multiple directions that have been pointed out and left as future work.

On the dynamic adaptation of language models based on dialogue information

We present an approach to adapt dynamically the language models (LMs) used by a speech recognizer that is part of a spoken dialogue system. We have developed a grammar generation strategy that automatically adapts the LMs using the semantic information that the user provides (represented as dialogue concepts), together with the information regarding the intentions of the speaker (inferred by the dialogue manager, and represented as dialogue goals). We carry out the adaptation as a linear interpolation between a background LM, and one or more of the LMs associated to the dialogue elements (concepts or goals) addressed by the user. The interpolation weights between those models are automatically estimated on each dialogue turn, using measures such as the posterior probabilities of concepts and goals, estimated as part of the inference procedure to determine the actions to be carried out. We propose two approaches to handle the LMs related to concepts and goals. Whereas in the first one we estimate a LM for each one of them, in the second one we apply several clustering strategies to group together those elements that share some common properties, and estimate a LM for each cluster. Our evaluation shows how the system can estimate a dynamic model adapted to each dialogue turn, which helps to improve the performance of the speech recognition (up to a 14.82% of relative improvement), which leads to an improvement in both the language understanding and the dialogue management tasks.

Deployment and validation of a smart system for screening of language disorders in primary care

Neuro-evolutive development from birth until the age of six years is a decisive factor in a child?s quality of life. Early detection of development disorders in early childhood can facilitate necessary diagnosis and/or treatment. Primary-care pediatricians play a key role in its detection as they can undertake the preventive and therapeutic actions requested to promote a child?s optimal development. However, the lack of time and little specific knowledge at primary-care avoid to applying continuous early-detection anomalies procedures. This research paper focuses on the deployment and evaluation of a smart system that enhances the screening of language disorders in primary care. Pediatricians get support to proceed with early referral of language disorders. The proposed model provides them with a decision-support tool for referral actions to trigger essential diagnostic and/or therapeutic actions for a comprehensive individual development. The research was conducted by starting from a sample of 60 cases of children with language disorders. Validation was carried out through two complementary steps: first, by including a team of seven experts from the fields of neonatology, pediatrics, neurology and language therapy, and, second, through the evaluation of 21 more previously diagnosed cases. The results obtained show that therapist positively accepted the system proposal in 18 cases (86%) and suggested system redesign for single referral to a speech therapist in three remaining cases.

Low-resource language recognition using a fusion of phoneme posteriorgram counts, acoustic and glottal-based i-vectors

This paper presents a description of our system for the Albayzin 2012 LRE competition. One of the main characteristics of this evaluation was the reduced number of available files for training the system, especially for the empty condition where no training data set was provided but only a development set. In addition, the whole database was created from online videos and around one third of the training data was labeled as noisy files. Our primary system was the fusion of three different i-vector based systems: one acoustic system based on MFCCs, a phonotactic system using trigrams of phone-posteriorgram counts, and another acoustic system based on RPLPs that improved robustness against noise. A contrastive system that included new features based on the glottal source was also presented. Official and postevaluation results for all the conditions using the proposed metrics for the evaluation and the Cavg metric are presented in the paper.

Speech features of the domain of the word in heterosexual twins during the second yeard of life

Speech is the major function, emergence and which development radically changes all course of formation of the identity of the child already in the early childhood. If language and speech development in solitary born children is investigated today quite well, at twin children this process practically is not studied. Our research was carried out for the purpose of studying of an originality of mastering by speech by heterosexual children of pair of twins within communicative and pragmatist approach (T.N. Ushakov,G. V. Chirkina). Application of this approach to the analysis of process of communication at twin children allowed us to allocate those peculiar receptions and means of communication which they functionally develop in a situation of pair of twins, as allows them to show the phenomena of the speech which are not meeting at solitary born contemporaries. In this work results of supervision and research of pair of heterosexual twins of the second year of the life, carried out by a technique developed by us under the scientific guide of G. V. Chirkina

Multilingual number transcription for text-to-speech conversion

This paper describes the text normalization module of a text to speech fully-trainable conversion system and its application to number transcription. The main target is to generate a language independent text normalization module, based on data instead of on expert rules. This paper proposes a general architecture based on statistical machine translation techniques. This proposal is composed of three main modules: a tokenizer for splitting the text input into a token graph, a phrase-based translation module for token translation, and a post-processing module for removing some tokens. This architecture has been evaluated for number transcription in several languages: English, Spanish and Romanian. Number transcription is an important aspect in the text normalization problem.

A supervised cooperative learning system for early detection of language disorders

The Quality of Life of a person may depend on early attention to his neurodevel-opment disorders in childhood. Identification of language disorders under the age of six years old can speed up required diagnosis and/or treatment processes. This paper details the enhancement of a Clinical Decision Support System (CDSS) aimed to assist pediatricians and language therapists at early identification and re-ferral of language disorders. The system helps to fine tune the Knowledge Base of Language Delays (KBLD) that was already developed and validated in clinical routine with 146 children. Medical experts supported the construction of Gades CDSS by getting scientific consensus from literature and fifteen years of regis-tered use cases of children with language disorders. The current research focuses on an innovative cooperative model that allows the evolution of the KBLD of Gades through the supervised evaluation of the CDSS learnings with experts¿ feedback. The deployment of the resulting system is being assessed under a mul-tidisciplinary team of seven experts from the fields of speech therapist, neonatol-ogy, pediatrics, and neurology.

Evaluating a web-based clinical decision support system for language disorders screening in a nursery school

Background: Early and effective identification of developmental disorders during childhood remains a critical task for the international community. The second highest prevalence of common developmental disorders in children are language delays, which are frequently the first symptoms of a possible disorder. Objective: This paper evaluates a Web-based Clinical Decision Support System (CDSS) whose aim is to enhance the screening of language disorders at a nursery school. The common lack of early diagnosis of language disorders led us to deploy an easy-to-use CDSS in order to evaluate its accuracy in early detection of language pathologies. This CDSS can be used by pediatricians to support the screening of language disorders in primary care. Methods: This paper details the evaluation results of the ?Gades? CDSS at a nursery school with 146 children, 12 educators, and 1 language therapist. The methodology embraces two consecutive phases. The first stage involves the observation of each child?s language abilities, carried out by the educators, to facilitate the evaluation of language acquisition level performed by a language therapist. Next, the same language therapist evaluates the reliability of the observed results. Results: The Gades CDSS was integrated to provide the language therapist with the required clinical information. The validation process showed a global 83.6% (122/146) success rate in language evaluation and a 7% (7/94) rate of non-accepted system decisions within the range of children from 0 to 3 years old. The system helped language therapists to identify new children with potential disorders who required further evaluation. This process will revalidate the CDSS output and allow the enhancement of early detection of language disorders in children. The system does need minor refinement, since the therapists disagreed with some questions from the CDSS knowledge base (KB) and suggested adding a few questions about speech production and pragmatic abilities. The refinement of the KB will address these issues and include the requested improvements, with the support of the experts who took part in the original KB development. Conclusions: This research demonstrated the benefit of a Web-based CDSS to monitor children?s neurodevelopment via the early detection of language delays at a nursery school. Current next steps focus on the design of a model that includes pseudo auto-learning capacity, supervised by experts.

Translating bus information into sign language for deaf people

This paper describes the application of language translation technologies for generating bus information in Spanish Sign Language (LSE: Lengua de Signos Española). In this work, two main systems have been developed: the first for translating text messages from information panels and the second for translating spoken Spanish into natural conversations at the information point of the bus company. Both systems are made up of a natural language translator (for converting a word sentence into a sequence of LSE signs), and a 3D avatar animation module (for playing back the signs). For the natural language translator, two technological approaches have been analyzed and integrated: an example-based strategy and a statistical translator. When translating spoken utterances, it is also necessary to incorporate a speech recognizer for decoding the spoken utterance into a word sequence, prior to the language translation module. This paper includes a detailed description of the field evaluation carried out in this domain. This evaluation has been carried out at the customer information office in Madrid involving both real bus company employees and deaf people. The evaluation includes objective measurements from the system and information from questionnaires. In the field evaluation, the whole translation presents an SER (Sign Error Rate) of less than 10% and a BLEU greater than 90%.

Dynamic topic-based adaptation of language models: a comparison between different approaches

This paper presents a dynamic LM adaptation based on the topic that has been identified on a speech segment. We use LSA and the given topic labels in the training dataset to obtain and use the topic models. We propose a dynamic language model adaptation to improve the recognition performance in "a two stages" AST system. The final stage makes use of the topic identification with two variants: the first on uses just the most probable topic and the other one depends on the relative distances of the topics that have been identified. We perform the adaptation of the LM as a linear interpolation between a background model and topic-based LM. The interpolation weight id dynamically adapted according to different parameters. The proposed method is evaluated on the Spanish partition of the EPPS speech database. We achieved a relative reduction in WER of 11.13% over the baseline system which uses a single blackground LM.

Extended phone log-likelihood ratio features and acoustic-based I-vectors for language recognition

This paper presents new techniques with relevant improvements added to the primary system presented by our group to the Albayzin 2012 LRE competition, where the use of any additional corpora for training or optimizing the models was forbidden. In this work, we present the incorporation of an additional phonotactic subsystem based on the use of phone log-likelihood ratio features (PLLR) extracted from different phonotactic recognizers that contributes to improve the accuracy of the system in a 21.4% in terms of Cavg (we also present results for the official metric during the evaluation, Fact). We will present how using these features at the phone state level provides significant improvements, when used together with dimensionality reduction techniques, especially PCA. We have also experimented with applying alternative SDC-like configurations on these PLLR features with additional improvements. Also, we will describe some modifications to the MFCC-based acoustic i-vector system which have also contributed to additional improvements. The final fused system outperformed the baseline in 27.4% in Cavg.

Language recognition using phonotactic-based shifted delta coefficients and multiple phone recognizers

A new language recognition technique based on the application of the philosophy of the Shifted Delta Coefficients (SDC) to phone log-likelihood ratio features (PLLR) is described. The new methodology allows the incorporation of long-span phonetic information at a frame-by-frame level while dealing with the temporal length of each phone unit. The proposed features are used to train an i-vector based system and tested on the Albayzin LRE 2012 dataset. The results show a relative improvement of 33.3% in Cavg in comparison with different state-of-the-art acoustic i-vector based systems. On the other hand, the integration of parallel phone ASR systems where each one is used to generate multiple PLLR coefficients which are stacked together and then projected into a reduced dimension are also presented. Finally, the paper shows how the incorporation of state information from the phone ASR contributes to provide additional improvements and how the fusion with the other acoustic and phonotactic systems provides an important improvement of 25.8% over the system presented during the competition.

GTH-UPM system for search on speech evaluation

This paper describes the GTH-UPM system for the Albayzin 2014 Search on Speech Evaluation. Teh evaluation task consists of searching a list of terms/queries in audio files. The GTH-UPM system we are presenting is based on a LVCSR (Large Vocabulary Continuous Speech Recognition) system. We have used MAVIR corpus and the Spanish partition of the EPPS (European Parliament Plenary Sessions) database for training both acoustic and language models. The main effort has been focused on lexicon preparation and text selection for the language model construction. The system makes use of different lexicon and language models depending on the task that is performed. For the best configuration of the system on the development set, we have obtained a FOM of 75.27 for the deyword spotting task.

Contributions to Speech Analytics based on Speech Recognition and Topic Identification

La última década ha sido testigo de importantes avances en el campo de la tecnología de reconocimiento de voz. Los sistemas comerciales existentes actualmente poseen la capacidad de reconocer habla continua de múltiples locutores, consiguiendo valores aceptables de error, y sin la necesidad de realizar procedimientos explícitos de adaptación. A pesar del buen momento que vive esta tecnología, el reconocimiento de voz dista de ser un problema resuelto. La mayoría de estos sistemas de reconocimiento se ajustan a dominios particulares y su eficacia depende de manera significativa, entre otros muchos aspectos, de la similitud que exista entre el modelo de lenguaje utilizado y la tarea específica para la cual se está empleando. Esta dependencia cobra aún más importancia en aquellos escenarios en los cuales las propiedades estadísticas del lenguaje varían a lo largo del tiempo, como por ejemplo, en dominios de aplicación que involucren habla espontánea y múltiples temáticas. En los últimos años se ha evidenciado un constante esfuerzo por mejorar los sistemas de reconocimiento para tales dominios. Esto se ha hecho, entre otros muchos enfoques, a través de técnicas automáticas de adaptación. Estas técnicas son aplicadas a sistemas ya existentes, dado que exportar el sistema a una nueva tarea o dominio puede requerir tiempo a la vez que resultar costoso. Las técnicas de adaptación requieren fuentes adicionales de información, y en este sentido, el lenguaje hablado puede aportar algunas de ellas. El habla no sólo transmite un mensaje, también transmite información acerca del contexto en el cual se desarrolla la comunicación hablada (e.g. acerca del tema sobre el cual se está hablando). Por tanto, cuando nos comunicamos a través del habla, es posible identificar los elementos del lenguaje que caracterizan el contexto, y al mismo tiempo, rastrear los cambios que ocurren en estos elementos a lo largo del tiempo. Esta información podría ser capturada y aprovechada por medio de técnicas de recuperación de información (information retrieval) y de aprendizaje de máquina (machine learning). Esto podría permitirnos, dentro del desarrollo de mejores sistemas automáticos de reconocimiento de voz, mejorar la adaptación de modelos del lenguaje a las condiciones del contexto, y por tanto, robustecer al sistema de reconocimiento en dominios con condiciones variables (tales como variaciones potenciales en el vocabulario, el estilo y la temática). En este sentido, la principal contribución de esta Tesis es la propuesta y evaluación de un marco de contextualización motivado por el análisis temático y basado en la adaptación dinámica y no supervisada de modelos de lenguaje para el robustecimiento de un sistema automático de reconocimiento de voz. Esta adaptación toma como base distintos enfoque de los sistemas mencionados (de recuperación de información y aprendizaje de máquina) mediante los cuales buscamos identificar las temáticas sobre las cuales se está hablando en una grabación de audio. Dicha identificación, por lo tanto, permite realizar una adaptación del modelo de lenguaje de acuerdo a las condiciones del contexto. El marco de contextualización propuesto se puede dividir en dos sistemas principales: un sistema de identificación de temática y un sistema de adaptación dinámica de modelos de lenguaje. Esta Tesis puede describirse en detalle desde la perspectiva de las contribuciones particulares realizadas en cada uno de los campos que componen el marco propuesto: _ En lo referente al sistema de identificación de temática, nos hemos enfocado en aportar mejoras a las técnicas de pre-procesamiento de documentos, asimismo en contribuir a la definición de criterios más robustos para la selección de index-terms. – La eficiencia de los sistemas basados tanto en técnicas de recuperación de información como en técnicas de aprendizaje de máquina, y específicamente de aquellos sistemas que particularizan en la tarea de identificación de temática, depende, en gran medida, de los mecanismos de preprocesamiento que se aplican a los documentos. Entre las múltiples operaciones que hacen parte de un esquema de preprocesamiento, la selección adecuada de los términos de indexado (index-terms) es crucial para establecer relaciones semánticas y conceptuales entre los términos y los documentos. Este proceso también puede verse afectado, o bien por una mala elección de stopwords, o bien por la falta de precisión en la definición de reglas de lematización. En este sentido, en este trabajo comparamos y evaluamos diferentes criterios para el preprocesamiento de los documentos, así como también distintas estrategias para la selección de los index-terms. Esto nos permite no sólo reducir el tamaño de la estructura de indexación, sino también mejorar el proceso de identificación de temática. – Uno de los aspectos más importantes en cuanto al rendimiento de los sistemas de identificación de temática es la asignación de diferentes pesos a los términos de acuerdo a su contribución al contenido del documento. En este trabajo evaluamos y proponemos enfoques alternativos a los esquemas tradicionales de ponderado de términos (tales como tf-idf ) que nos permitan mejorar la especificidad de los términos, así como también discriminar mejor las temáticas de los documentos. _ Respecto a la adaptación dinámica de modelos de lenguaje, hemos dividimos el proceso de contextualización en varios pasos. – Para la generación de modelos de lenguaje basados en temática, proponemos dos tipos de enfoques: un enfoque supervisado y un enfoque no supervisado. En el primero de ellos nos basamos en las etiquetas de temática que originalmente acompañan a los documentos del corpus que empleamos. A partir de estas, agrupamos los documentos que forman parte de la misma temática y generamos modelos de lenguaje a partir de dichos grupos. Sin embargo, uno de los objetivos que se persigue en esta Tesis es evaluar si el uso de estas etiquetas para la generación de modelos es óptimo en términos del rendimiento del reconocedor. Por esta razón, nosotros proponemos un segundo enfoque, un enfoque no supervisado, en el cual el objetivo es agrupar, automáticamente, los documentos en clusters temáticos, basándonos en la similaridad semántica existente entre los documentos. Por medio de enfoques de agrupamiento conseguimos mejorar la cohesión conceptual y semántica en cada uno de los clusters, lo que a su vez nos permitió refinar los modelos de lenguaje basados en temática y mejorar el rendimiento del sistema de reconocimiento. – Desarrollamos diversas estrategias para generar un modelo de lenguaje dependiente del contexto. Nuestro objetivo es que este modelo refleje el contexto semántico del habla, i.e. las temáticas más relevantes que se están discutiendo. Este modelo es generado por medio de la interpolación lineal entre aquellos modelos de lenguaje basados en temática que estén relacionados con las temáticas más relevantes. La estimación de los pesos de interpolación está basada principalmente en el resultado del proceso de identificación de temática. – Finalmente, proponemos una metodología para la adaptación dinámica de un modelo de lenguaje general. El proceso de adaptación tiene en cuenta no sólo al modelo dependiente del contexto sino también a la información entregada por el proceso de identificación de temática. El esquema usado para la adaptación es una interpolación lineal entre el modelo general y el modelo dependiente de contexto. Estudiamos también diferentes enfoques para determinar los pesos de interpolación entre ambos modelos. Una vez definida la base teórica de nuestro marco de contextualización, proponemos su aplicación dentro de un sistema automático de reconocimiento de voz. Para esto, nos enfocamos en dos aspectos: la contextualización de los modelos de lenguaje empleados por el sistema y la incorporación de información semántica en el proceso de adaptación basado en temática. En esta Tesis proponemos un marco experimental basado en una arquitectura de reconocimiento en ‘dos etapas’. En la primera etapa, empleamos sistemas basados en técnicas de recuperación de información y aprendizaje de máquina para identificar las temáticas sobre las cuales se habla en una transcripción de un segmento de audio. Esta transcripción es generada por el sistema de reconocimiento empleando un modelo de lenguaje general. De acuerdo con la relevancia de las temáticas que han sido identificadas, se lleva a cabo la adaptación dinámica del modelo de lenguaje. En la segunda etapa de la arquitectura de reconocimiento, usamos este modelo adaptado para realizar de nuevo el reconocimiento del segmento de audio. Para determinar los beneficios del marco de trabajo propuesto, llevamos a cabo la evaluación de cada uno de los sistemas principales previamente mencionados. Esta evaluación es realizada sobre discursos en el dominio de la política usando la base de datos EPPS (European Parliamentary Plenary Sessions - Sesiones Plenarias del Parlamento Europeo) del proyecto europeo TC-STAR. Analizamos distintas métricas acerca del rendimiento de los sistemas y evaluamos las mejoras propuestas con respecto a los sistemas de referencia. ABSTRACT The last decade has witnessed major advances in speech recognition technology. Today’s commercial systems are able to recognize continuous speech from numerous speakers, with acceptable levels of error and without the need for an explicit adaptation procedure. Despite this progress, speech recognition is far from being a solved problem. Most of these systems are adjusted to a particular domain and their efficacy depends significantly, among many other aspects, on the similarity between the language model used and the task that is being addressed. This dependence is even more important in scenarios where the statistical properties of the language fluctuates throughout the time, for example, in application domains involving spontaneous and multitopic speech. Over the last years there has been an increasing effort in enhancing the speech recognition systems for such domains. This has been done, among other approaches, by means of techniques of automatic adaptation. These techniques are applied to the existing systems, specially since exporting the system to a new task or domain may be both time-consuming and expensive. Adaptation techniques require additional sources of information, and the spoken language could provide some of them. It must be considered that speech not only conveys a message, it also provides information on the context in which the spoken communication takes place (e.g. on the subject on which it is being talked about). Therefore, when we communicate through speech, it could be feasible to identify the elements of the language that characterize the context, and at the same time, to track the changes that occur in those elements over time. This information can be extracted and exploited through techniques of information retrieval and machine learning. This allows us, within the development of more robust speech recognition systems, to enhance the adaptation of language models to the conditions of the context, thus strengthening the recognition system for domains under changing conditions (such as potential variations in vocabulary, style and topic). In this sense, the main contribution of this Thesis is the proposal and evaluation of a framework of topic-motivated contextualization based on the dynamic and non-supervised adaptation of language models for the enhancement of an automatic speech recognition system. This adaptation is based on an combined approach (from the perspective of both information retrieval and machine learning fields) whereby we identify the topics that are being discussed in an audio recording. The topic identification, therefore, enables the system to perform an adaptation of the language model according to the contextual conditions. The proposed framework can be divided in two major systems: a topic identification system and a dynamic language model adaptation system. This Thesis can be outlined from the perspective of the particular contributions made in each of the fields that composes the proposed framework: _ Regarding the topic identification system, we have focused on the enhancement of the document preprocessing techniques in addition to contributing in the definition of more robust criteria for the selection of index-terms. – Within both information retrieval and machine learning based approaches, the efficiency of topic identification systems, depends, to a large extent, on the mechanisms of preprocessing applied to the documents. Among the many operations that encloses the preprocessing procedures, an adequate selection of index-terms is critical to establish conceptual and semantic relationships between terms and documents. This process might also be weakened by a poor choice of stopwords or lack of precision in defining stemming rules. In this regard we compare and evaluate different criteria for preprocessing the documents, as well as for improving the selection of the index-terms. This allows us to not only reduce the size of the indexing structure but also to strengthen the topic identification process. – One of the most crucial aspects, in relation to the performance of topic identification systems, is to assign different weights to different terms depending on their contribution to the content of the document. In this sense we evaluate and propose alternative approaches to traditional weighting schemes (such as tf-idf ) that allow us to improve the specificity of terms, and to better identify the topics that are related to documents. _ Regarding the dynamic language model adaptation, we divide the contextualization process into different steps. – We propose supervised and unsupervised approaches for the generation of topic-based language models. The first of them is intended to generate topic-based language models by grouping the documents, in the training set, according to the original topic labels of the corpus. Nevertheless, a goal of this Thesis is to evaluate whether or not the use of these labels to generate language models is optimal in terms of recognition accuracy. For this reason, we propose a second approach, an unsupervised one, in which the objective is to group the data in the training set into automatic topic clusters based on the semantic similarity between the documents. By means of clustering approaches we expect to obtain a more cohesive association of the documents that are related by similar concepts, thus improving the coverage of the topic-based language models and enhancing the performance of the recognition system. – We develop various strategies in order to create a context-dependent language model. Our aim is that this model reflects the semantic context of the current utterance, i.e. the most relevant topics that are being discussed. This model is generated by means of a linear interpolation between the topic-based language models related to the most relevant topics. The estimation of the interpolation weights is based mainly on the outcome of the topic identification process. – Finally, we propose a methodology for the dynamic adaptation of a background language model. The adaptation process takes into account the context-dependent model as well as the information provided by the topic identification process. The scheme used for the adaptation is a linear interpolation between the background model and the context-dependent one. We also study different approaches to determine the interpolation weights used in this adaptation scheme. Once we defined the basis of our topic-motivated contextualization framework, we propose its application into an automatic speech recognition system. We focus on two aspects: the contextualization of the language models used by the system, and the incorporation of semantic-related information into a topic-based adaptation process. To achieve this, we propose an experimental framework based in ‘a two stages’ recognition architecture. In the first stage of the architecture, Information Retrieval and Machine Learning techniques are used to identify the topics in a transcription of an audio segment. This transcription is generated by the recognition system using a background language model. According to the confidence on the topics that have been identified, the dynamic language model adaptation is carried out. In the second stage of the recognition architecture, an adapted language model is used to re-decode the utterance. To test the benefits of the proposed framework, we carry out the evaluation of each of the major systems aforementioned. The evaluation is conducted on speeches of political domain using the EPPS (European Parliamentary Plenary Sessions) database from the European TC-STAR project. We analyse several performance metrics that allow us to compare the improvements of the proposed systems against the baseline ones.