Deployment of human-machine dialogue systems.

The deployment of systems for human-to-machine communication by voice requires overcoming a variety of obstacles that affect the speech-processing technologies. Problems encountered in the field might include variation in speaking style, acoustic noise, ambiguity of language, or confusion on the part of the speaker. The diversity of these practical problems encountered in the "real world" leads to the perceived gap between laboratory and "real-world" performance. To answer the question "What applications can speech technology support today?" the concept of the "degree of difficulty" of an application is introduced. The degree of difficulty depends not only on the demands placed on the speech recognition and speech synthesis technologies but also on the expectations of the user of the system. Experience has shown that deployment of effective speech communication systems requires an iterative process. This paper discusses general deployment principles, which are illustrated by several examples of human-machine communication systems.

What does voice-processing technology support today?

This paper describes the state of the art in applications of voice-processing technologies. In the first part, technologies concerning the implementation of speech recognition and synthesis algorithms are described. Hardware technologies such as microprocessors and DSPs (digital signal processors) are discussed. Software development environment, which is a key technology in developing applications software, ranging from DSP software to support software also is described. In the second part, the state of the art of algorithms from the standpoint of applications is discussed. Several issues concerning evaluation of speech recognition/synthesis algorithms are covered, as well as issues concerning the robustness of algorithms in adverse conditions.

The future of voice-processing technology in the world of computers and communications.

This talk, which was the keynote address of the NAS Colloquium on Human-Machine Communication by Voice, discusses the past, present, and future of human-machine communications, especially speech recognition and speech synthesis. Progress in these technologies is reviewed in the context of the general progress in computer and communications technologies.

Reconhecimento de monossílabos em idosos: análise do nível de apresentação da fala

A perda auditiva no idoso acarreta em dificuldade na percepção da fala. O teste comumente utilizado na logoaudiometria é a pesquisa do índice de reconhecimento de fala máximo (IR-Max) em uma única intensidade de apresentação da fala. Entretanto, o procedimento mais adequado seria a realização do teste em diversas intensidades, visto que o índice de acerto depende da intensidade da fala no momento do teste e está relacionado com o grau e configuração da perda auditiva. A imprecisão na obtenção do IR-Max poderá gerar uma hipótese diagnóstica errônea e o insucesso no processo de intervenção na perda auditiva. Objetivo: Verificar a interferência do nível de apresentação da fala, no teste de reconhecimento de fala, em idosos com perda auditiva sensorioneural com diferentes configurações audiométricas. Métodos: Participaram 64 idosos, 120 orelhas (61 do gênero feminino e 59 do gênero masculino), idade entre 60 e 88 anos, divididos em grupos: G1- composto por 23 orelhas com configuração horizontal, G2- 55 orelhas com configuração descendente, G3- 42 orelhas com configuração abrupta. Os critérios de inclusão foram: perda auditiva sensorioneural de grau leve a severo, não usuário de aparelho de amplificação sonora individual (AASI), ou com tempo de uso inferior a dois meses, e ausência de alterações cognitivas. Foram realizados os seguintes procedimentos: pesquisas do limiar de reconhecimento de fala (LRF), do índice de reconhecimento de fala (IRF) em diversas intensidades e do nível de máximo conforto (MCL) e desconforto (UCL) para a fala. Para tal, foram utilizadas listas com 11 monossílabos, para diminuir a duração do teste. A análise estatística foi composta pelo teste Análise de Variância (ANOVA) e teste de Tukey. Resultados: A configuração descendente foi a de maior ocorrência. Indivíduos com configuração horizontal apresentaram índice médio de acerto mais elevado de reconhecimento de fala. Ao considerar o total avaliado, 27,27% dos indivíduos com configuração horizontal revelaram o IR-Max no MCL, assim como 38,18% com configuração descendente e 26,19% com configuração abrupta. O IR-Max foi encontrado no UCL, em 40,90% dos indivíduos com configuração horizontal, 45,45% com configuração descendente e 28,20% com configuração abrupta. Respectivamente, o maior e o menor índice médio de acerto foram encontrados em: G1- 30 e 40 dBNS; G2- 50 e 10 dBNS; G3- 45 e 10 dBNS. Não há uma única intensidade de fala a ser utilizada em todos os tipos de configurações audiométricas, entretanto, os níveis de sensação que identificaram os maiores índices médios de acerto foram: G1- 20 a 30 dBNS, G2- 20 a 50 dBNS; G3- 45 dBNS. O MCL e o UCL-5 dB para a fala não foram eficazes para determinar o IR-Max. Conclusões: O nível de apresentação teve influência no desempenho no reconhecimento de fala para monossílabos em idosos com perda auditiva sensorioneural em todas as configurações audiométricas. A perda auditiva de grau moderado e a configuração audiométrica descendente foram mais frequentes nessa população, seguida da abrupta e horizontal.

Designing Regularizers and Architectures for Recurrent Neural Networks

Cette thèse contribue a la recherche vers l'intelligence artificielle en utilisant des méthodes connexionnistes. Les réseaux de neurones récurrents sont un ensemble de modèles séquentiels de plus en plus populaires capable en principe d'apprendre des algorithmes arbitraires. Ces modèles effectuent un apprentissage en profondeur, un type d'apprentissage machine. Sa généralité et son succès empirique en font un sujet intéressant pour la recherche et un outil prometteur pour la création de l'intelligence artificielle plus générale. Le premier chapitre de cette thèse donne un bref aperçu des sujets de fonds: l'intelligence artificielle, l'apprentissage machine, l'apprentissage en profondeur et les réseaux de neurones récurrents. Les trois chapitres suivants couvrent ces sujets de manière de plus en plus spécifiques. Enfin, nous présentons quelques contributions apportées aux réseaux de neurones récurrents. Le chapitre \ref{arxiv1} présente nos travaux de régularisation des réseaux de neurones récurrents. La régularisation vise à améliorer la capacité de généralisation du modèle, et joue un role clé dans la performance de plusieurs applications des réseaux de neurones récurrents, en particulier en reconnaissance vocale. Notre approche donne l'état de l'art sur TIMIT, un benchmark standard pour cette tâche. Le chapitre \ref{cpgp} présente une seconde ligne de travail, toujours en cours, qui explore une nouvelle architecture pour les réseaux de neurones récurrents. Les réseaux de neurones récurrents maintiennent un état caché qui représente leurs observations antérieures. L'idée de ce travail est de coder certaines dynamiques abstraites dans l'état caché, donnant au réseau une manière naturelle d'encoder des tendances cohérentes de l'état de son environnement. Notre travail est fondé sur un modèle existant; nous décrivons ce travail et nos contributions avec notamment une expérience préliminaire.

Google Translate e Microsoft Translator - Valutazione di due applicazioni per la traduzione automatica del parlato e analisi di una tecnologia in evoluzione.

This work focuses on Machine Translation (MT) and Speech-to-Speech Translation, two emerging technologies that allow users to automatically translate written and spoken texts. The first part of this work provides a theoretical framework for the evaluation of Google Translate and Microsoft Translator, which is at the core of this study. Chapter one focuses on Machine Translation, providing a definition of this technology and glimpses of its history. In this chapter we will also learn how MT works, who uses it, for what purpose, what its pros and cons are, and how machine translation quality can be defined and assessed. Chapter two deals with Speech-to-Speech Translation by focusing on its history, characteristics and operation, potential uses and limits deriving from the intrinsic difficulty of translating spoken language. After describing the future prospects for SST, the final part of this chapter focuses on the quality assessment of Speech-to-Speech Translation applications. The last part of this dissertation describes the evaluation test carried out on Google Translate and Microsoft Translator, two mobile translation apps also providing a Speech-to-Speech Translation service. Chapter three illustrates the objectives, the research questions, the participants, the methodology and the elaboration of the questionnaires used to collect data. The collected data and the results of the evaluation of the automatic speech recognition subsystem and the language translation subsystem are presented in chapter four and finally analysed and compared in chapter five, which provides a general description of the performance of the evaluated apps and possible explanations for each set of results. In the final part of this work suggestions are made for future research and reflections on the usability and usefulness of the evaluated translation apps are provided.

Designing Regularizers and Architectures for Recurrent Neural Networks

Cette thèse contribue a la recherche vers l'intelligence artificielle en utilisant des méthodes connexionnistes. Les réseaux de neurones récurrents sont un ensemble de modèles séquentiels de plus en plus populaires capable en principe d'apprendre des algorithmes arbitraires. Ces modèles effectuent un apprentissage en profondeur, un type d'apprentissage machine. Sa généralité et son succès empirique en font un sujet intéressant pour la recherche et un outil prometteur pour la création de l'intelligence artificielle plus générale. Le premier chapitre de cette thèse donne un bref aperçu des sujets de fonds: l'intelligence artificielle, l'apprentissage machine, l'apprentissage en profondeur et les réseaux de neurones récurrents. Les trois chapitres suivants couvrent ces sujets de manière de plus en plus spécifiques. Enfin, nous présentons quelques contributions apportées aux réseaux de neurones récurrents. Le chapitre \ref{arxiv1} présente nos travaux de régularisation des réseaux de neurones récurrents. La régularisation vise à améliorer la capacité de généralisation du modèle, et joue un role clé dans la performance de plusieurs applications des réseaux de neurones récurrents, en particulier en reconnaissance vocale. Notre approche donne l'état de l'art sur TIMIT, un benchmark standard pour cette tâche. Le chapitre \ref{cpgp} présente une seconde ligne de travail, toujours en cours, qui explore une nouvelle architecture pour les réseaux de neurones récurrents. Les réseaux de neurones récurrents maintiennent un état caché qui représente leurs observations antérieures. L'idée de ce travail est de coder certaines dynamiques abstraites dans l'état caché, donnant au réseau une manière naturelle d'encoder des tendances cohérentes de l'état de son environnement. Notre travail est fondé sur un modèle existant; nous décrivons ce travail et nos contributions avec notamment une expérience préliminaire.

New Techniques in Deep Representation Learning

Thesis (Ph.D.)--University of Washington, 2016-06

Submodular Optimization and Data Processing

Thesis (Ph.D.)--University of Washington, 2016-06

Harmonic Encoding in Cochlear Implants

Thesis (Master's)--University of Washington, 2016-06

Proceedings of the Eighth Australian and New Zealand Intelligent Information Systems Conference

These are the full proceedings of the conference.

Neural networks for time-varying data

This paper reviews some basic issues and methods involved in using neural networks to respond in a desired fashion to a temporally-varying environment. Some popular network models and training methods are introduced. A speech recognition example is then used to illustrate the central difficulty of temporal data processing: learning to notice and remember relevant contextual information. Feedforward network methods are applicable to cases where this problem is not severe. The application of these methods are explained and applications are discussed in the areas of pure mathematics, chemical and physical systems, and economic systems. A more powerful but less practical algorithm for temporal problems, the moving targets algorithm, is sketched and discussed. For completeness, a few remarks are made on reinforcement learning.

Perception of concurrent sentences with harmonic or frequency-shifted voiced excitation:performance of human listeners and of computational models based on autocorrelation

Keyword identification in one of two simultaneous sentences is improved when the sentences differ in F0, particularly when they are almost continuously voiced. Sentences of this kind were recorded, monotonised using PSOLA, and re-synthesised to give a range of harmonic ?F0s (0, 1, 3, and 10 semitones). They were additionally re-synthesised by LPC with the LPC residual frequency shifted by 25% of F0, to give excitation with inharmonic but regularly spaced components. Perceptual identification of frequency-shifted sentences showed a similar large improvement with nominal ?F0 as seen for harmonic sentences, although overall performance was about 10% poorer. We compared performance with that of two autocorrelation-based computational models comprising four stages: (i) peripheral frequency selectivity and half-wave rectification; (ii) within-channel periodicity extraction; (iii) identification of the two major peaks in the summary autocorrelation function (SACF); (iv) a template-based approach to speech recognition using dynamic time warping. One model sampled the correlogram at the target-F0 period and performed spectral matching; the other deselected channels dominated by the interferer and performed matching on the short-lag portion of the residual SACF. Both models reproduced the monotonic increase observed in human performance with increasing ?F0 for the harmonic stimuli, but not for the frequency-shifted stimuli. A revised version of the spectral-matching model, which groups patterns of periodicity that lie on a curve in the frequency-delay plane, showed a closer match to the perceptual data for frequency-shifted sentences. The results extend the range of phenomena originally attributed to harmonic processing to grouping by common spectral pattern.

Verbal Dialogue versus Written Dialogue

Modern technology has moved on and completely changed the way that people can use the telephone or mobile to dialogue with information held on computers. Well developed “written speech analysis” does not work with “verbal speech”. The main purpose of our article is, firstly, to highlights the problems and, secondly, to shows the possible ways to solve these problems.

Studio e sviluppo di un framework per il riconoscimento vocale nell'ambito di sistemi Hands-Free

Negli ultimi anni, l'avanzamento incredibilmente rapido della tecnologia ha portato allo sviluppo e alla diffusione di dispositivi elettronici portatili aventi dimensioni estremamente ridotte e, allo stesso tempo, capacità computazionali molto notevoli. Più nello specifico, una particolare categoria di dispositivi, attualmente in forte sviluppo, che ha già fatto la propria comparsa sul mercato mondiale è sicuramente la categoria dei dispositivi Wearable. Come suggerisce il nome, questi sono progettati per essere letteralmente indossati, pensati per fornire continuo supporto, in diversi ambiti, a chi li utilizza. Se per interagire con essi l’utente non deve ricorrere obbligatoriamente all'utilizzo delle mani, allora si parla di dispositivi Wearable Hands Free. Questi sono generalmente in grado di percepire e catture l’input dell'utente seguendo tecniche e metodologie diverse, non basate sul tatto. Una di queste è sicuramente quella che prevede di modellare l’input dell’utente stesso attraverso la sua voce, appoggiandosi alla disciplina dell’ASR (Automatic Speech Recognition), che si occupa della traduzione del linguaggio parlato in testo, mediante l’utilizzo di dispositivi computerizzati. Si giunge quindi all’obiettivo della tesi, che è quello di sviluppare un framework, utilizzabile nell’ambito dei dispositivi Wearable, che fornisca un servizio di riconoscimento vocale appoggiandosi ad uno già esistente, in modo che presenti un certo livello di efficienza e facilità di utilizzo. Più in generale, in questo documento si punta a fornire una descrizione approfondita di quelli che sono i dispositivi Wearable e Wearable Hands-Free, definendone caratteristiche, criticità e ambiti di utilizzo. Inoltre, l’intento è quello di illustrare i principi di funzionamento dell’Automatic Speech Recognition per passare poi ad analisi, progettazione e sviluppo del framework appena citato.