| | --- INTERSPEECH 2014 - SINGAPORE
--- September 14-18, 2014
--- http://www.INTERSPEECH2014.org
The INTERSPEECH 2014 Organising Committee is pleased to announce
the following 8 tutorials presented by distinguished speakers
at the conference and will be offered on Sunday, 14 September 2014.
All Tutorials will be of three (3) hours duration and require
an additional registration fee (separate from the conference registration fee).
• Non-speech acoustic event detection and classification
• Contribution of MRI to Exploring and Modeling Speech Production
• Computational Models for Audiovisual Emotion Perception
• The Art and Science of Speech Feature Engineering
• Recent Advances in Speaker Diarization
• Multimodal Speech Recognition with the AusTalk 3D Audio-Visual Corpus
• Semantic Web and Linked Big Data Resources for Spoken Language Processing
• Speech and Audio for Multimedia Semantics
----------------------------------------------------------------------------------------------------
ISCSLP Tutorials @ INTERSPEECH 2014
----------------------------------------------------------------------------------------------------
Additionally, the ISCSLP 2014 Organising Committee welcomes
the INTERSPEECH 2014 delegates to join the 4 ISCSLP tutorials
which will be offered on Saturday, 13 September 2014.
• Adaptation Techniques for Statistical Speech Recognition
• Emotion and Mental State Recognition: Features, Models, System Applications and Beyond
• Unsupervised Speech and Language Processing via Topic Models
• Deep Learning for Speech Generation and Synthesis
More information available at: http://www.interspeech2014.org/public.php?page=tutorial.html
----------------------------------------------------------------------------------------------------
Tutorials Description
----------------------------------------------------------------------------------------------------
T1: Non-speech acoustic event detection and classification
The research in audio signal processing has been dominated by speech research,
but most of the sounds in our real-life environments are actually non-speech
events such as cars passing by, wind, warning beeps, and animal sounds.
These acoustic events contain much information about the environment and physical
events that take place in it, enabling novel application areas such as safety,
health monitoring and investigation of biodiversity. But while recent years
have seen wide-spread adoption of applications such as speech recognition and
song recognition, generic computer audition is still in its infancy.
Non-speech acoustic events have several fundamental differences to speech,
but many of the core algorithms used by speech researchers can be leveraged
for generic audio analysis. The tutorial is a comprehensive review of the field
of acoustic event detection as it currently stands. The goal of the tutorial is
foster interest in the community, highlight the challenges and opportunities
and provide a starting point for new researchers. We will discuss what acoustic
event detection entails, the commonalities differences with speech processing,
such as the large variation in sounds and the possible overlap with other sounds.
We will then discuss basic experimental and algorithm design, including descriptions
of available databases and machine learning methods. We will then discuss more
advanced topics such as methods to deal with temporally overlapping sounds and
modelling the relations between sounds. We will finish with a discussion of
avenues for future research.
Organizers: Tuomas Virtanen and Jort F. Gemmeke
T2: Contribution of MRI to Exploring and Modeling Speech Production
Magnetic resonance imaging (MRI) provides us a magic vision to look into
the human body in various ways not only with static imaging but also with
motion imaging. MRI has been a powerful technique for speech research to
study finer anatomy of the speech organs or to visualize true vocal tracts
in three dimensions. Inherent problems of slow image acquisition for speech
tasks or insufficient signal-to-noise ratio for microscopic observation have
been the cost for researchers to search for task-specific imaging techniques.
The recent advances of the 3-Tesla technology suggest more practical solutions
to broader applications of MRI by overcoming previous technical limitations.
In this joint tutorial in two parts, we summarize our previous effort to accumulate
scientific knowledge with MRI and to advance speech modeling studies for future
development. Part I, given by Kiyoshi Honda, introduces how to visualize the
speech organs and vocal tracts by presenting techniques and data for finer static
imaging, synchronized motion imaging, surface marker tracking, real-time imaging,
and vocal-tract mechanical modeling. Part 2, presented by Jianwu Dang, focuses on
applications of MRI for phonetics of Mandarin vowels, acoustics of the vocal tracts
with side branches, analysis and simulation in search of talker characteristics,
physiological modeling of the articulatory system, and motor control paradigm
for speech articulation.
Organizers: Kiyoshi HONDA and Jianwu DANG
T3: Computational Models for Audiovisual Emotion Perception
In this tutorial we will explore engineering approaches to understanding human
emotion perception, focusing both on modeling and application. We will highlight
both current and historical trends in emotion perception modeling, focusing on
both psychological and engineering-driven theories of perception
(statistical analyses, data-driven computational modeling, and implicit sensing).
The importance of this topic can be appreciated from both an engineering viewpoint,
any system that either models human behavior or interacts with human partners must
understand emotion perception as it fundamentally underlies and modulates our
communication, or from a psychological perspective, emotion perception is also used
in the diagnosis of many mental health conditions and is tracked in therapeutic
interventions. Research in emotion perception seeks to identify models that describe
the felt sense of ‘typical’ emotion expression – i.e., an observer/evaluator’s attribution
of the emotional state of the speaker. This felt sense is a function of the methods through
which individuals integrate the presented multimodal emotional information.
We will cover psychological theories of emotion, engineering models of emotion,
and experimental approaches to measure emotion. We will demonstrate how these modeling
strategies can be used as a component of emotion classification frameworks and how
they can be used to inform the design of emotional behaviors.
Organizers: Emily Mower Provost and Carlos Busso
T4: The Art and Science of Speech Feature Engineering
With significant advances in mobile technology and audio sensing devices,
there is a fundamental need to describe vast amounts of audio data in terms
of well representative lower dimensional descriptors for efficient automatic
processing. The extraction of these signal representations, also called features,
constitutes the first step in processing a speech signal. The art and science of
feature engineering relates to addressing the two inherent challenges - extracting
sufficient information from the speech signal for the task at hand and suppressing
the unwanted redundancies for computational efficiency and robustness.
The area of speech feature extraction combines a wide variety of disciplines like
signal processing, machine learning, psychophysics, information theory, linguistics and physiology.
It has a rich history spanning more than five decades and has seen tremendous advances
in the last few years. This has propelled the transition of the speech technology from
controlled environments to millions of end user applications.
In this tutorial, we review the evolution of speech feature processing methods,
summarize the recent advances of the last two decades and provide insights into the
future of feature engineering. This will include the discussions on the spectral
representation methods developed in the past, human auditory motivated techniques
for robust speech processing, data driven unsupervised features like ivectors and
recent advances in deep neural network based techniques. With experimental results,
we will also illustrate the impact of these features for various state-of-the-art
speech processing systems. The future of speech signal processing will need to address
various robustness issues in complex acoustic environments while being able
to derive useful information from big data.
Organizers: Sriram Ganapathy and Samuel Thomas
T5: Recent Advances in Speaker Diarization
The tutorial will start with an introduction to speaker diarization giving a general
overview of the subject. Afterwards, we will cover the basic background including
feature extraction, and common modeling techniques such as GMMs and HMMs.
Then, we will discuss the first processing step usually done in speaker diarization
which is voice activity detection. We will consequently describe the classic approaches
for speaker diarization which are widely used today. We will then introduce state-of-the-art
techniques in speaker recognition required to understand modern speaker diarization techniques.
Following, we will describe approaches for speaker diarization using advanced representation
methods (supervectors, speaker factors, i-vectors) and we will describe supervised and
unsupervised learning techniques used for speaker diarization. We will also discuss issues
such as coping with unknown number of speakers, detecting and dealing with overlapping speech,
diarization confidence estimation, and online speaker diarization. Finally we will discuss
two recent works: exploiting a-prioiri acoustic information (such as processing a meeting
when some of the participants are known in advanced to the system, and training data is available for them),
The second recent work is modeling speaker-turn dynamics. If time permits, we will also discuss concepts
such as multi-modal diarization and using TDOA (time difference of arrival) for diarization of meetings.
Organizers: Hagai Aronowitz
T6: Multimodal Speech Recognition with the AusTalk 3D Audio-Visual Corpus
This tutorial will provide attendees a brief overview of 3D based AVSR research.
In this tutorial, attendees will learn how to use the newly developed 3D based audio
visual data corpus we derived from the AusTalk corpus (https://austalk.edu.au/)
for audio-visual speech/speaker recognition. In addition, we also plan to introduce
some results using this newly developed 3D audio-visual data corpus, which show that
there is a significant speech accuracy increase by integrating both depth-level and grey-level
visual features. In the first part of the tutorial, we will review some recent works published
in the last decade, so that attendees can obtain an overview of the fundamental concepts
and challenges in this field. In the second part of the tutorial, we will briefly describe
the recording protocol and contents of the 3D data corpus, and show attendees how to use
this corpus for their own research. In the third part of this tutorial, we will present our
results using the 3D data corpus. The experimental results show that, compared with the
conventional AVSR based on the audio and grey-level visual features, the integration of grey
and depth visual information can boost the AVSR accuracy significantly. Moreover,
we will also experimentally explain why adding depth information can benefit the standard AVSR systems.
Eventually, through our tutorial, we hope we can inspire more researchers in the community
to contribute to this exciting research.
Organizers: Roberto Togneri, Mohammed Bennamoun and Chao (Luke) Sui
T7: Semantic Web and Linked Big Data Resources for Spoken Language Processing
State-of-the-art statistical spoken language processing typically requires
significant manual effort to construct domain-specific schemas (ontologies)
as well as manual effort to annotate training data against these schemas.
At the same time, a recent surge of activity and progress on semantic web-related
concepts from the large search-engine companies represents a potential alternative
to the manually intensive design of spoken language processing systems.
Standards such as schema.org have been established for schemas (ontologies) that
webmasters can use to semantically and uniformly markup their web pages.
Search engines like Bing, Google, and Yandex have adopted these standards and are
leveraging them to create semantic search engines at the scale of the web.
As a result, the open linked data resources and semantic graphs covering various
domains (such as Freebase [3]) have grown massively every year and contains far more
information than any single resource anywhere on the Web. Furthermore, these resources
contain links to text data (such as Wikipedia pages) related to the knowledge in the graph.
Recently, several studies on speech language processing started exploiting these massive
linked data resources for language modeling and spoken language understanding.
This tutorial will include a brief introduction to the semantic web and the linked
data structure, available resources, and querying languages.
An overview of related work on information extraction and language processing will
be presented, where the main focus will be on methods for learning spoken language
understanding models from these resources.
Organizers: Dilek Hakkani-Tür and Larry Heck
T8: Speech and Audio for Multimedia Semantics
Internet media sharing sites and the one-click upload capability of smartphones
are producing a deluge of multimedia content. While visual features are often dominant
in such material, acoustic and speech information in particular often complements it.
By facilitating access to large amounts of data, the text-based Internet gave a huge
boost to the field of natural language processing. The vast amount of consumer-produced
video becoming available now will do the same for video processing, eventually enabling
semantic understanding of multimedia material, with implications for human computer interaction, robotics, etc.
Large-scale multi-modal analysis of audio-visual material is now central to a number of
multi-site research projects around the world. While each of these have slightly different
targets, they are facing largely the same challenges: how to robustly and efficiently process
large amounts of data, how to represent and then fuse information across modalities,
how to train classifiers and segmenters on unlabeled data, how to include human feedback, etc.
In this tutorial, we will present the state of the art in large-scale video, speech,
and non-speech audio processing, and show how these approaches are being applied to tasks
such as content based video retrieval (CBVR) and multimedia event detection (MED).
We will introduce the most important tools and techniques, and show how the combination of
information across modalities can be used to induce semantics on multimedia material
through ranking of information and fusion. Finally, we will discuss opportunities
for research that the INTERSPEECH community specifically will find interesting and fertile.
Organizers: Florian Metze and Koichi Shinoda
----------------------------------------------------------------------------------------------------
ISCSLP Tutorials @ INTERSPEECH 2014 Description
----------------------------------------------------------------------------------------------------
ISCSLP-T1: Adaptation Techniques for Statistical Speech Recognition
Adaptation is a technique to make better use of existing models for test data
from new acoustic or linguistic conditions. It is an important and challenging
research area of statistical speech recognition. This tutorial gives a systematic
review of fundamental theories as well as introduction of state-of-the-art adaptation
techniques. It includes both acoustic and language model adaptation. Following a simple example
of acoustic model adaptation, basic concepts, procedures and categories of adaptation will
be introduced. Then, a number of advanced adaptation techniques will be discussed,
such as discriminative adaptation, Deep Neural Network adaptation, adaptive training,
relationship to noise robustness etc. After the detailed review of acoustic model adaptation,
an introduction of language model adaptation, such as topic adaptation will also be given.
The whole tutorial is then summarised and future research direction will be discussed.
Organizers: Kai Yu
ISCSLP-T2: Emotion and Mental State Recognition: Features, Models, System Applications and Beyond
Emotion recognition is the ability to identify what you are feeling from moment
to moment and to understand the connection between your feelings and your expressions.
In today’s world, human-computer interaction (HCI) interface undoubtedly plays an
important role in our daily life. Toward harmonious HCI interfaces, automated analysis
and recognition of human emotion has attracted increasing attention from researchers
in multidisciplinary research fields. A specific area of current interest that also has key
implications for HCI is the estimation of cognitive load (mental workload), research into
which is still at an early stage. Technologies for processing daily activities including speech,
text and music have expanded the interaction modalities between humans and computer-supported
communicational artifacts.
In this tutorial, we will present theoretical and practical work offering new and broad views
of the latest research in emotional awareness from audio and speech. We discuss several parts
spanning a variety of theoretical background and applications ranging from salient emotional features,
emotional-cognitive models, compensation methods for variability due to speaker and linguistic content,
to machine learning approaches applicable to emotion recognition. In each topic, we will review
the state of the art by introducing current methods and presenting several applications.
In particular, the application to cognitive load estimation will be discussed,
from its psychophysiological origins to system design considerations. Eventually,
technologies developed in different areas will be combined for future applications,
so in addition to a survey of future research challenges,
we will envision a few scenarios in which affective computing can make a difference.
Organizers: Chung-Hsien Wu, Hsin-Min Wang, Julien Epps and Vidhyasaharan Sethu
ISCSLP-T3: Unsupervised Speech and Language Processing via Topic Models
In this tutorial, we will present state-of-art machine learning approaches
for speech and language processing with highlight on the unsupervised methods
for structural learning from the unlabeled sequential patterns. In general,
speech and language processing involves extensive knowledge of statistical models.
We require designing a flexible, scalable and robust system to meet heterogeneous
and nonstationary environments in the era of big data. This tutorial starts from an
introduction of unsupervised speech and language processing based on factor analysis
and independent component analysis. The unsupervised learning is generalized to a latent
variable model which is known as the topic model. The evolution of topic models from
latent semantic analysis to hierarchical Dirichlet process, from non-Bayesian parametric
models to Bayesian nonparametric models, and from single-layer model to hierarchical
tree model shall be surveyed in an organized fashion. The inference approaches based on
variational Bayesian and Gibbs sampling are introduced. We will also present several
case studies on topic modeling for speech and language applications including language model,
document model, retrieval model, segmentation model and summarization model.
At last, we will point out new trends of topic models for speech and language processing.
Organizers: Jen-Tzung Chien
ISCSLP-T4: Deep Learning for Speech Generation and Synthesis
Deep learning, which can represent high-level abstractions in data with an architecture of
multiple non-linear transformation, has made a huge impact on automatic speech recognition (ASR)
research, products and services. However, deep learning for speech generation and synthesis
(i.e., text-to-speech), which is an inverse process of speech recognition (i.e., speech-to-text),
has not generated the similar momentum as it is for ASR yet. Recently, motivated by the success
of Deep Neural Networks in speech recognition, some neural network based research attempts have
been tried successfully on improving the performance of statistical parametric based
speech generation/synthesis. In this tutorial, we focus on deep learning approaches to the
problems in speech generation and synthesis, especially on Text-to-Speech (TTS) synthesis and voice conversion.
First, we give a review for the current main stream of statistical parametric based speech generation
and synthesis, or the GMM-HMM based speech synthesis and GMM-based voice conversion with emphasis
on analyzing the major factors responsible for the quality problems in the GMM-based voice
synthesis/conversion and the intrinsic limitations of a decision-tree based, contextual state
clustering and state-based statistical distribution modeling. We then present the latest deep
learning algorithms for feature parameter trajectory generation, in contrast to deep learning for
recognition or classification. We cover common technologies in Deep Neural Network (DNN) and improved
DNN: Mixture Density Networks (MDN), Recurrent Neural Networks (RNN) with Bidirectional Long Short
Term Memory (BLSTM) and Conditional RBM (CRBM). Finally, we share our research insights and hand-on
experience on building speech generation and synthesis systems based upon deep learning algorithms.
Organizers: Yao Qian and Frank K. Soong |
