Voice conversion with pre-trained representations for audio anonymization

The recording and processing of voice data raises increasing privacy concerns for users and service providers. One way to address these issues is to move processing on the edge device closer to the recording so that potentially identifiable information is not transmitted over the internet. However, this is often not possible due to hardware limitations. An interesting alternative is the development of voice anonymization techniques that remove individual speakers characteristics while preserving linguistic and acoustic information in the data. In this work, a state-of-the-art approach to sequence-to-sequence speech conversion, ini- tially based on x-vectors and bottleneck features for automatic speech recognition, is explored to disentangle the two acoustic information using different pre-trained speech and speakers representation. Furthermore, different strategies for selecting target speech representations are analyzed. Results on public datasets in terms of equal error rate and word error rate show that good privacy is achieved with limited impact on converted speech quality relative to the original method.

Automatic Understanding of ATC Speech: Study of Prospectives and Field Experiments for Several Controller Positions

Although there has been a lot of interest in recognizing and understanding air traffic control (ATC) speech, none of the published works have obtained detailed field data results. We have developed a system able to identify the language spoken and recognize and understand sentences in both Spanish and English. We also present field results for several in-tower controller positions. To the best of our knowledge, this is the first time that field ATC speech (not simulated) is captured, processed, and analyzed. The use of stochastic grammars allows variations in the standard phraseology that appear in field data. The robust understanding algorithm developed has 95% concept accuracy from ATC text input. It also allows changes in the presentation order of the concepts and the correction of errors created by the speech recognition engine improving it by 17% and 25%, respectively, absolute in the percentage of fully correctly understood sentences for English and Spanish in relation to the percentages of fully correctly recognized sentences. The analysis of errors due to the spontaneity of the speech and its comparison to read speech is also carried out. A 96% word accuracy for read speech is reduced to 86% word accuracy for field ATC data for Spanish for the "clearances" task confirming that field data is needed to estimate the performance of a system. A literature review and a critical discussion on the possibilities of speech recognition and understanding technology applied to ATC speech are also given.

Improving automatic detection of obstructive sleep apnea through nonlinear analysis of sustained speech

We present a novel approach for the detection of severe obstructive sleep apnea (OSA) based on patients' voices introducing nonlinear measures to describe sustained speech dynamics. Nonlinear features were combined with state-of-the-art speech recognition systems using statistical modeling techniques (Gaussian mixture models, GMMs) over cepstral parameterization (MFCC) for both continuous and sustained speech. Tests were performed on a database including speech records from both severe OSA and control speakers. A 10 % relative reduction in classification error was obtained for sustained speech when combining MFCC-GMM and nonlinear features, and 33 % when fusing nonlinear features with both sustained and continuous MFCC-GMM. Accuracy reached 88.5 % allowing the system to be used in OSA early detection. Tests showed that nonlinear features and MFCCs are lightly correlated on sustained speech, but uncorrelated on continuous speech. Results also suggest the existence of nonlinear effects in OSA patients' voices, which should be found in continuous speech.

The development of an automatic recognition system for earmark and earprint comparisons

The value of earmarks as an efficient means of personal identification is still subject to debate. It has been argued that the field is lacking a firm systematic and structured data basis to help practitioners to form their conclusions. Typically, there is a paucity of research guiding as to the selectivity of the features used in the comparison process between an earmark and reference earprints taken from an individual. This study proposes a system for the automatic comparison of earprints and earmarks, operating without any manual extraction of key-points or manual annotations. For each donor, a model is created using multiple reference prints, hence capturing the donor within source variability. For each comparison between a mark and a model, images are automatically aligned and a proximity score, based on a normalized 2D correlation coefficient, is calculated. Appropriate use of this score allows deriving a likelihood ratio that can be explored under known state of affairs (both in cases where it is known that the mark has been left by the donor that gave the model and conversely in cases when it is established that the mark originates from a different source). To assess the system performance, a first dataset containing 1229 donors elaborated during the FearID research project was used. Based on these data, for mark-to-print comparisons, the system performed with an equal error rate (EER) of 2.3% and about 88% of marks are found in the first 3 positions of a hitlist. When performing print-to-print transactions, results show an equal error rate of 0.5%. The system was then tested using real-case data obtained from police forces.

Robust and fast vowel recognition using optimum-path forest

The applications of Automatic Vowel Recognition (AVR), which is a sub-part of fundamental importance in most of the speech processing systems, vary from automatic interpretation of spoken language to biometrics. State-of-the-art systems for AVR are based on traditional machine learning models such as Artificial Neural Networks (ANNs) and Support Vector Machines (SVMs), however, such classifiers can not deal with efficiency and effectiveness at the same time, existing a gap to be explored when real-time processing is required. In this work, we present an algorithm for AVR based on the Optimum-Path Forest (OPF), which is an emergent pattern recognition technique recently introduced in literature. Adopting a supervised training procedure and using speech tags from two public datasets, we observed that OPF has outperformed ANNs, SVMs, plus other classifiers, in terms of training time and accuracy. ©2010 IEEE.

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.

Toward the ultimate synthesis/recognition system.

This paper predicts speech synthesis, speech recognition, and speaker recognition technology for the year 2001, and it describes the most important research problems to be solved in order to arrive at these ultimate synthesis and recognition systems. The problems for speech synthesis include natural and intelligible voice production, prosody control based on meaning, capability of controlling synthesized voice quality and choosing individual speaking style, multilingual and multidialectal synthesis, choice of application-oriented speaking styles, capability of adding emotion, and synthesis from concepts. The problems for speech recognition include robust recognition against speech variations, adaptation/normalization to variations due to environmental conditions and speakers, automatic knowledge acquisition for acoustic and linguistic modeling, spontaneous speech recognition, naturalness and ease of human-machine interaction, and recognition of emotion. The problems for speaker recognition are similar to those for speech recognition. The research topics related to all these techniques include the use of articulatory and perceptual constraints and evaluation methods for measuring the quality of technology and systems.

A Cognitive Science Reasoning in Recognition of Emotions in Audio-Visual Speech

In this report we summarize the state-of-the-art of speech emotion recognition from the signal processing point of view. On the bases of multi-corporal experiments with machine-learning classifiers, the observation is made that existing approaches for supervised machine learning lead to database dependent classifiers which can not be applied for multi-language speech emotion recognition without additional training because they discriminate the emotion classes following the used training language. As there are experimental results showing that Humans can perform language independent categorisation, we made a parallel between machine recognition and the cognitive process and tried to discover the sources of these divergent results. The analysis suggests that the main difference is that the speech perception allows extraction of language independent features although language dependent features are incorporated in all levels of the speech signal and play as a strong discriminative function in human perception. Based on several results in related domains, we have suggested that in addition, the cognitive process of emotion-recognition is based on categorisation, assisted by some hierarchical structure of the emotional categories, existing in the cognitive space of all humans. We propose a strategy for developing language independent machine emotion recognition, related to the identification of language independent speech features and the use of additional information from visual (expression) features.

Speech intelligibility with and without noise in individuals exposed to electronic music

Audiometry is the main way with which hearing is evaluated, because it is a universal and standardized test. Speech tests are difficult to standardize due to the variables involved, their performance in the presence of competitive noise is of great importance. Aim: To characterize speech intelligibility in silence and in competitive noise from individuals exposed to electronically amplified music. Material and Method: It was performed with 20 university students who presented normal hearing thresholds. The speech recognition rate (SRR) was performed after fourteen hours of sound rest after the exposure to electronically amplified music and once again after sound rest, being studied in three stages: without competitive noise, in the presence of Babble-type competitive noise, in monotic listening, in signal/ noise ratio of + 5 dB and with the signal/ noise ratio of 5 dB. Results: There was greater damage in the SRR after exposure to the music and with competitive noise, and as the signal/ noise ratio decreases, the performance of individuals in the test also decreased. Conclusion: The inclusion of competitive noise in the speech tests in the audiological routine is important, because it represents the real disadvantage experienced by individuals in daily listening.

Analysis and recognition of similar environmental sounds

Dissertação apresentada na Faculdade de Ciências e Tecnologia da Universidade Nova de Lisboa para obtenção do grau de Mestre em Engenharia Informática

Adaptive modeling and high quality spectral estimation for speech enhancement

In this work an adaptive modeling and spectral estimation scheme based on a dual Discrete Kalman Filtering (DKF) is proposed for speech enhancement. Both speech and noise signals are modeled by an autoregressive structure which provides an underlying time frame dependency and improves time-frequency resolution. The model parameters are arranged to obtain a combined state-space model and are also used to calculate instantaneous power spectral density estimates. The speech enhancement is performed by a dual discrete Kalman filter that simultaneously gives estimates for the models and the signals. This approach is particularly useful as a pre-processing module for parametric based speech recognition systems that rely on spectral time dependent models. The system performance has been evaluated by a set of human listeners and by spectral distances. In both cases the use of this pre-processing module has led to improved results.

Speech as the basic interface for assistive technology

Speech interfaces for Assistive Technologies are not common and are usually replaced by others. The market they are targeting is not considered attractive and speech technologies are still not well spread. Industry still thinks they present some performance risks, especially Speech Recognition systems. As speech is the most elemental and natural way for communication, it has strong potential for enhancing inclusion and quality of life for broader groups of users with special needs, such as people with cerebral palsy and elderly staying at their homes. This work is a position paper in which the authors argue for the need to make speech become the basic interface in assistive technologies. Among the main arguments, we can state: speech is the easiest way to interact with machines; there is a growing market for embedded speech in assistive technologies, since the number of disabled and elderly people is expanding; speech technology is already mature to be used but needs adaptation to people with special needs; there is still a lot of R&D to be done in this area, especially when thinking about the Portuguese market. The main challenges are presented and future directions are proposed.

Aplicacions de reconeixement de la parla: sistemes de respeaking III

En aquest projecte es fa una introducció als reconeixedors de la parla, el seu funcionament i la seva base matemàtica. Un cop tots els conceptes han quedat clars, es mostra el mètode de creació que hem seguit per obtenir el nostre propi reconeixedor de la parla, utilitzant les eines HTK, en català. S’avaluen les seves virtuts i els seus defectes a través de diferents proves realitzades als seus components. A més a més, el projecte arrodoneix la feina implementant un sistema de dictat automàtic que explota el reconeixedor de la parla utilitzant Julius.

Pattern recognition of sleep in rodents using piezoelectric signals generated by gross body movements.

Current research on sleep using experimental animals is limited by the expense and time-consuming nature of traditional EEG/EMG recordings. We present here an alternative, noninvasive approach utilizing piezoelectric films configured as highly sensitive motion detectors. These film strips attached to the floor of the rodent cage produce an electrical output in direct proportion to the distortion of the material. During sleep, movement associated with breathing is the predominant gross body movement and, thus, output from the piezoelectric transducer provided an accurate respiratory trace during sleep. During wake, respiratory movements are masked by other motor activities. An automatic pattern recognition system was developed to identify periods of sleep and wake using the piezoelectric generated signal. Due to the complex and highly variable waveforms that result from subtle postural adjustments in the animals, traditional signal analysis techniques were not sufficient for accurate classification of sleep versus wake. Therefore, a novel pattern recognition algorithm was developed that successfully distinguished sleep from wake in approximately 95% of all epochs. This algorithm may have general utility for a variety of signals in biomedical and engineering applications. This automated system for monitoring sleep is noninvasive, inexpensive, and may be useful for large-scale sleep studies including genetic approaches towards understanding sleep and sleep disorders, and the rapid screening of the efficacy of sleep or wake promoting drugs.

Towards a lightweight adaptative multicomponential affective recognition system

As part of the Affective Computing research field, the development of automatic affective recognition systems can enhance human-computer interactions by allowing the creation of interfaces that react to the user's emotional state. To that end, this Master Thesis brings affect recognition to nowadays most used human computer interface, mobile devices, by developing a facial expression recognition system able to perform detection under the difficult conditions of viewing angle and illumination that entails the interaction with a mobile device. Moreover, this Master Thesis proposes to combine emotional features detected from expression with contextual information of the current situation, to infer a complex and extensive emotional state of the user. Thus, a cognitive computational model of emotion is defined that provides a multicomponential affective state of the user through the integration of the detected emotional features into appraisal processes. In order to account for individual differences in the emotional experience, these processes can be adapted to the culture and personality of the user.