A Satisfaction-based Model for Affect Recognition from Conversational Features in Spoken Dialog Systems

Detecting user affect automatically during real-time conversation is the main challenge towards our greater aim of infusing social intelligence into a natural-language mixed-initiative High-Fidelity (Hi-Fi) audio control spoken dialog agent. In recent years, studies on affect detection from voice have moved on to using realistic, non-acted data, which is subtler. However, it is more challenging to perceive subtler emotions and this is demonstrated in tasks such as labelling and machine prediction. This paper attempts to address part of this challenge by considering the role of user satisfaction ratings and also conversational/dialog features in discriminating contentment and frustration, two types of emotions that are known to be prevalent within spoken human-computer interaction. However, given the laboratory constraints, users might be positively biased when rating the system, indirectly making the reliability of the satisfaction data questionable. Machine learning experiments were conducted on two datasets, users and annotators, which were then compared in order to assess the reliability of these datasets. Our results indicated that standard classifiers were significantly more successful in discriminating the abovementioned emotions and their intensities (reflected by user satisfaction ratings) from annotator data than from user data. These results corroborated that: first, satisfaction data could be used directly as an alternative target variable to model affect, and that they could be predicted exclusively by dialog features. Second, these were only true when trying to predict the abovementioned emotions using annotator?s data, suggesting that user bias does exist in a laboratory-led evaluation.

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.

Towards glottal source controllability in expressive speech synthesis

In order to obtain more human like sounding humanmachine interfaces we must first be able to give them expressive capabilities in the way of emotional and stylistic features so as to closely adequate them to the intended task. If we want to replicate those features it is not enough to merely replicate the prosodic information of fundamental frequency and speaking rhythm. The proposed additional layer is the modification of the glottal model, for which we make use of the GlottHMM parameters. This paper analyzes the viability of such an approach by verifying that the expressive nuances are captured by the aforementioned features, obtaining 95% recognition rates on styled speaking and 82% on emotional speech. Then we evaluate the effect of speaker bias and recording environment on the source modeling in order to quantify possible problems when analyzing multi-speaker databases. Finally we propose a speaking styles separation for Spanish based on prosodic features and check its perceptual significance.

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.

Evaluation of a transplantation algorithm for expressive speech synthesis

When designing human-machine interfaces it is important to consider not only the bare bones functionality but also the ease of use and accessibility it provides. When talking about voice-based inter- faces, it has been proven that imbuing expressiveness into the synthetic voices increases signi?cantly its perceived naturalness, which in the end is very helpful when building user friendly interfaces. This paper proposes an adaptation based expressiveness transplantation system capable of copying the emotions of a source speaker into any desired target speaker with just a few minutes of read speech and without requiring the record- ing of additional expressive data. This system was evaluated through a perceptual test for 3 speakers showing up to an average of 52% emotion recognition rates relative to the natural voice recognition rates, while at the same time keeping good scores in similarity and naturality.

Classical vs. biometric features in the 2013 speaker recognition evaluation in mobile environments

MFCC coefficients extracted from the power spectral density of speech as a whole, seems to have become the de facto standard in the area of speaker recognition, as demonstrated by its use in almost all systems submitted to the 2013 Speaker Recognition Evaluation (SRE) in Mobile Environment [1], thus relegating to background this component of the recognition systems. However, in this article we will show that selecting the adequate speaker characterization system is as important as the selection of the classifier. To accomplish this we will compare the recognition rates achieved by different recognition systems that relies on the same classifier (GMM-UBM) but connected with different feature extraction systems (based on both classical and biometric parameters). As a result we will show that a gender dependent biometric parameterization with a simple recognition system based on GMM- UBM paradigm provides very competitive or even better recognition rates when compared to more complex classification systems based on classical features

Speech Signals Feature Extraction Model for a Speaker’s Gender and Age Identification System

Durante el proceso de producción de voz, los factores anatómicos, fisiológicos o psicosociales del individuo modifican los órganos resonadores, imprimiendo en la voz características particulares. Los sistemas ASR tratan de encontrar los matices característicos de una voz y asociarlos a un individuo o grupo. La edad y sexo de un hablante son factores intrínsecos que están presentes en la voz. Este trabajo intenta diferenciar esas características, aislarlas y usarlas para detectar el género y la edad de un hablante. Para dicho fin, se ha realizado el estudio y análisis de las características basadas en el pulso glótico y el tracto vocal, evitando usar técnicas clásicas (como pitch y sus derivados) debido a las restricciones propias de dichas técnicas. Los resultados finales de nuestro estudio alcanzan casi un 100% en reconocimiento de género mientras en la tarea de reconocimiento de edad el reconocimiento se encuentra alrededor del 80%. Parece ser que la voz queda afectada por el género del hablante y las hormonas, aunque no se aprecie en la audición. ABSTRACT Particular elements of the voice are printed during the speech production process and are related to anatomical and physiological factors of the phonatory system or psychosocial factors acquired by the speaker. ASR systems attempt to find those peculiar nuances of a voice and associate them to an individual or a group. Age and gender are inherent factors to the speaker which may be represented in voice. This work attempts to differentiate those characteristics, isolate them and use them to detect speaker’s gender and age. Features based on glottal pulse and vocal tract are studied and analyzed in order to achieve good results in both tasks. Classical methodologies (such as pitch and derivates) are avoided since the requirements of those techniques may be too restrictive. The final scores achieve almost 100% in gender recognition whereas in age recognition those scores are around 80%. Factors related to the gender and hormones seem to affect the voice although they are not audible.

New methodology for the integration of biometric features in speaker recognition systems applied to security environments

La cuestión principal abordada en esta tesis doctoral es la mejora de los sistemas biométricos de reconocimiento de personas a partir de la voz, proponiendo el uso de una nueva parametrización, que hemos denominado parametrización biométrica extendida dependiente de género (GDEBP en sus siglas en inglés). No se propone una ruptura completa respecto a los parámetros clásicos sino una nueva forma de utilizarlos y complementarlos. En concreto, proponemos el uso de parámetros diferentes dependiendo del género del locutor, ya que como es bien sabido, la voz masculina y femenina presentan características diferentes que deberán modelarse, por tanto, de diferente manera. Además complementamos los parámetros clásicos utilizados (MFFC extraídos de la señal de voz), con un nuevo conjunto de parámetros extraídos a partir de la deconstrucción de la señal de voz en sus componentes de fuente glótica (más relacionada con el proceso y órganos de fonación y por tanto con características físicas del locutor) y de tracto vocal (más relacionada con la articulación acústica y por tanto con el mensaje emitido). Para verificar la validez de esta propuesta se plantean diversos escenarios, utilizando diferentes bases de datos, para validar que la GDEBP permite generar una descripción más precisa de los locutores que los parámetros MFCC clásicos independientes del género. En concreto se plantean diferentes escenarios de identificación sobre texto restringido y texto independiente utilizando las bases de datos de HESPERIA y ALBAYZIN. El trabajo también se completa con la participación en dos competiciones internacionales de reconocimiento de locutor, NIST SRE (2010 y 2012) y MOBIO 2013. En el primer caso debido a la naturaleza de las bases de datos utilizadas se obtuvieron resultados cercanos al estado del arte, mientras que en el segundo de los casos el sistema presentado obtuvo la mejor tasa de reconocimiento para locutores femeninos. A pesar de que el objetivo principal de esta tesis no es el estudio de sistemas de clasificación, sí ha sido necesario analizar el rendimiento de diferentes sistemas de clasificación, para ver el rendimiento de la parametrización propuesta. En concreto, se ha abordado el uso de sistemas de reconocimiento basados en el paradigma GMM-UBM, supervectores e i-vectors. Los resultados que se presentan confirman que la utilización de características que permitan describir los locutores de manera más precisa es en cierto modo más importante que la elección del sistema de clasificación utilizado por el sistema. En este sentido la parametrización propuesta supone un paso adelante en la mejora de los sistemas de reconocimiento biométrico de personas por la voz, ya que incluso con sistemas de clasificación relativamente simples se consiguen tasas de reconocimiento realmente competitivas. ABSTRACT The main question addressed in this thesis is the improvement of automatic speaker recognition systems, by the introduction of a new front-end module that we have called Gender Dependent Extended Biometric Parameterisation (GDEBP). This front-end do not constitute a complete break with respect to classical parameterisation techniques used in speaker recognition but a new way to obtain these parameters while introducing some complementary ones. Specifically, we propose a gender-dependent parameterisation, since as it is well known male and female voices have different characteristic, and therefore the use of different parameters to model these distinguishing characteristics should provide a better characterisation of speakers. Additionally, we propose the introduction of a new set of biometric parameters extracted from the components which result from the deconstruction of the voice into its glottal source estimate (close related to the phonation process and the involved organs, and therefore the physical characteristics of the speaker) and vocal tract estimate (close related to acoustic articulation and therefore to the spoken message). These biometric parameters constitute a complement to the classical MFCC extracted from the power spectral density of speech as a whole. In order to check the validity of this proposal we establish different practical scenarios, using different databases, so we can conclude that a GDEBP generates a more accurate description of speakers than classical approaches based on gender-independent MFCC. Specifically, we propose scenarios based on text-constrain and text-independent test using HESPERIA and ALBAYZIN databases. This work is also completed with the participation in two international speaker recognition evaluations: NIST SRE (2010 and 2012) and MOBIO 2013, with diverse results. In the first case, due to the nature of the NIST databases, we obtain results closed to state-of-the-art although confirming our hypothesis, whereas in the MOBIO SRE we obtain the best simple system performance for female speakers. Although the study of classification systems is beyond the scope of this thesis, we found it necessary to analise the performance of different classification systems, in order to verify the effect of them on the propose parameterisation. In particular, we have addressed the use of speaker recognition systems based on the GMM-UBM paradigm, supervectors and i-vectors. The presented results confirm that the selection of a set of parameters that allows for a more accurate description of the speakers is as important as the selection of the classification method used by the biometric system. In this sense, the proposed parameterisation constitutes a step forward in improving speaker recognition systems, since even when using relatively simple classification systems, really competitive recognition rates are achieved.

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.

Acoustic and neural bases for innate recognition of song

In behavior reminiscent of the responsiveness of human infants to speech, young songbirds innately recognize and prefer to learn the songs of their own species. The acoustic and physiological bases for innate recognition were investigated in fledgling white-crowned sparrows lacking song experience. A behavioral test revealed that the complete conspecific song was not essential for innate recognition: songs composed of single white-crowned sparrow phrases and songs played in reverse elicited vocal responses as strongly as did normal song. In all cases, these responses surpassed those to other species’ songs. Although auditory neurons in the song nucleus HVc and the underlying neostriatum of fledglings did not prefer conspecific song over foreign song, some neurons responded strongly to particular phrase types characteristic of white-crowned sparrows and, thus, could contribute to innate song recognition.

The neural development and organization of letter recognition: Evidence from functional neuroimaging, computational modeling, and behavioral studies

Although much of the brain’s functional organization is genetically predetermined, it appears that some noninnate functions can come to depend on dedicated and segregated neural tissue. In this paper, we describe a series of experiments that have investigated the neural development and organization of one such noninnate function: letter recognition. Functional neuroimaging demonstrates that letter and digit recognition depend on different neural substrates in some literate adults. How could the processing of two stimulus categories that are distinguished solely by cultural conventions become segregated in the brain? One possibility is that correlation-based learning in the brain leads to a spatial organization in cortex that reflects the temporal and spatial clustering of letters with letters in the environment. Simulations confirm that environmental co-occurrence does indeed lead to spatial localization in a neural network that uses correlation-based learning. Furthermore, behavioral studies confirm one critical prediction of this co-occurrence hypothesis, namely, that subjects exposed to a visual environment in which letters and digits occur together rather than separately (postal workers who process letters and digits together in Canadian postal codes) do indeed show less behavioral evidence for segregated letter and digit processing.

Research in speech communication.

Advances in digital speech processing are now supporting application and deployment of a variety of speech technologies for human/machine communication. In fact, new businesses are rapidly forming about these technologies. But these capabilities are of little use unless society can afford them. Happily, explosive advances in microelectronics over the past two decades have assured affordable access to this sophistication as well as to the underlying computing technology. The research challenges in speech processing remain in the traditionally identified areas of recognition, synthesis, and coding. These three areas have typically been addressed individually, often with significant isolation among the efforts. But they are all facets of the same fundamental issue--how to represent and quantify the information in the speech signal. This implies deeper understanding of the physics of speech production, the constraints that the conventions of language impose, and the mechanism for information processing in the auditory system. In ongoing research, therefore, we seek more accurate models of speech generation, better computational formulations of language, and realistic perceptual guides for speech processing--along with ways to coalesce the fundamental issues of recognition, synthesis, and coding. Successful solution will yield the long-sought dictation machine, high-quality synthesis from text, and the ultimate in low bit-rate transmission of speech. It will also open the door to language-translating telephony, where the synthetic foreign translation can be in the voice of the originating talker.

Towards Closed-loop Deep Brain Stimulation: Behavior Recognition from Human STN

Deep brain stimulation (DBS) provides significant therapeutic benefit for movement disorders such as Parkinson’s disease (PD). Current DBS devices lack real-time feedback (thus are open loop) and stimulation parameters are adjusted during scheduled visits with a clinician. A closed-loop DBS system may reduce power consumption and side effects by adjusting stimulation parameters based on patient’s behavior. Thus behavior detection is a major step in designing such systems. Various physiological signals can be used to recognize the behaviors. Subthalamic Nucleus (STN) Local field Potential (LFP) is a great candidate signal for the neural feedback, because it can be recorded from the stimulation lead and does not require additional sensors. This thesis proposes novel detection and classification techniques for behavior recognition based on deep brain LFP. Behavior detection from such signals is the vital step in developing the next generation of closed-loop DBS devices. LFP recordings from 13 subjects are utilized in this study to design and evaluate our method. Recordings were performed during the surgery and the subjects were asked to perform various behavioral tasks. Various techniques are used understand how the behaviors modulate the STN. One method studies the time-frequency patterns in the STN LFP during the tasks. Another method measures the temporal inter-hemispheric connectivity of the STN as well as the connectivity between STN and Pre-frontal Cortex (PFC). Experimental results demonstrate that different behaviors create different m odulation patterns in STN and it’s connectivity. We use these patterns as features to classify behaviors. A method for single trial recognition of the patient’s current task is proposed. This method uses wavelet coefficients as features and support vector machine (SVM) as the classifier for recognition of a selection of behaviors: speech, motor, and random. The proposed method is 82.4% accurate for the binary classification and 73.2% for classifying three tasks. As the next step, a practical behavior detection method which asynchronously detects behaviors is proposed. This method does not use any priori knowledge of behavior onsets and is capable of asynchronously detect the finger movements of PD patients. Our study indicates that there is a motor-modulated inter-hemispheric connectivity between LFP signals recorded bilaterally from STN. We utilize a non-linear regression method to measure this inter-hemispheric connectivity and to detect the finger movements. Our experimental results using STN LFP recorded from eight patients with PD demonstrate this is a promising approach for behavior detection and developing novel closed-loop DBS systems.

Auditory-visual speech integration by prelinguistic infants: Perception of an emergent consonant in the McGurk effect

The McGurk effect, in which auditory [ba] dubbed onto [go] lip movements is perceived as da or tha, was employed in a real-time task to investigate auditory-visual speech perception in prelingual infants. Experiments 1A and 1B established the validity of real-time dubbing for producing the effect. In Experiment 2, 4(1)/(2)-month-olds were tested in a habituation-test paradigm, in which 2 an auditory-visual stimulus was presented contingent upon visual fixation of a live face. The experimental group was habituated to a McGurk stimulus (auditory [ba] visual [ga]), and the control group to matching auditory-visual [ba]. Each group was then presented with three auditory-only test trials, [ba], [da], and [deltaa] (as in then). Visual-fixation durations in test trials showed that the experimental group treated the emergent percept in the McGurk effect, [da] or [deltaa], as familiar (even though they had not heard these sounds previously) and [ba] as novel. For control group infants [da] and [deltaa] were no more familiar than [ba]. These results are consistent with infants'perception of the McGurk effect, and support the conclusion that prelinguistic infants integrate auditory and visual speech information. (C) 2004 Wiley Periodicals, Inc.