Wake-Up-Word Technology and it - PowerPoint PPT Presentation

Loading...

PPT – Wake-Up-Word Technology and it PowerPoint presentation | free to download - id: 7919c3-NDUyN



Loading


The Adobe Flash plugin is needed to view this content

Get the plugin now

View by Category
About This Presentation
Title:

Wake-Up-Word Technology and it

Description:

Wake-Up-Word Technology and it s Applications Dr. Veton K puska Wake-Up-Word (WUW) Speech Recognition Voice Only Activated Air-Traffic Control of Unmanned Aircraft. – PowerPoint PPT presentation

Number of Views:5
Avg rating:3.0/5.0
Slides: 28
Provided by: Vet98
Learn more at: http://my.fit.edu
Category:

less

Write a Comment
User Comments (0)
Transcript and Presenter's Notes

Title: Wake-Up-Word Technology and it


1
Wake-Up-Word Technology and its Applications
  • Dr. Veton Këpuska
  • Wake-Up-Word (WUW) Speech Recognition Voice Only
    Activated Air-Traffic Control of Unmanned
    Aircraft.
  • Surveillance for suspicious activity, intrusion,
    and breach with a proven covert seismic
    technology known as Seismic Arrays for Local
    Situation Awareness (SALSA) with a proven speech
    recognition technology known as Wake-Up Word (WUW)

2
Spontaneous Human-Machine Interface
  • Architecture of WUW driven voice only interface
    system

See Previous Figure

See Previous Figure

Lowest Score of an OOV Sample

Lowest Score of an OOV Sample



DavidClark
Signal Enhancement
WUW
Headset
Algorithm
Recognizer
External
Speaker
3
Covert Perimeter Monitoring, Detection and Cueing
  • Seismic Arrays for Local Situation Awareness
    (SALSA) with a proven Wake-Up Word (WUW) speech
    recognition technology SALSAWUW

4
Past Projects
  • SAR-LAB Speech Processing, Analysis and
    Recognition
  • Processing Analyzing Signals
  • Digital Signal Processing
  • Recognition of Patterns
  • Pattern Recognition

5
Machine Learning
  • Dr. Georgios Anagnostopoulos

6
An Intro to Clustering
  • Clustering refers to grouping data points into
    natural / meaningful groupings. These
    groupings are called clusters.
  • What constitutes a natural grouping is
    sometimes very subjective See example below with
    2D data

1 group?
Group 1
Group 2
3 groups?
2 groups?
Obvious natural grouping
Not so obvious natural grouping
7
An Intro to Clustering
  • Nevertheless, clustering plays an important role
    in Machine Learning (ML) and other disciplines.
  • For example, it is reasonable to assume (at
    least, in principle) that two groups of money
    borrowers exist responsible and
    irresponsible. Financial institutions like
    banks, which need to make decisions whether they
    are going to lend money or not to a given
    customer, have to base their decision on the
    customers credit history (data points).
    Therefore it may make sense to group customers
    into 2 groups based on their credit history.
  • However, a more fine-grained grouping with 3
    clusters, such as responsible, occasionally
    responsible and irresponsible, may also be
    desirable. Notice the subjectivity involved in
    clustering depending on the particular
    application

8
An Intro to Clustering
  • When data points are high-dimensional (e.g.
    credit history of a consumer that consists of
    many observations), the data cannot be visualized
    so that they are grouped by an individual via
    visual inspection. Therefore, an automated
    clustering procedure is called for.
  • Also, the number of potential groupings of N data
    points into K non-empty clusters is given by the
    Stirling number of the 2nd kind S(N,K). Note that
    S(21, 2) gt 106 , which means that 21 can be
    grouped into two clusters in more than a million
    ways!!! There can be a huge number of possibly
    reasonable groupings, whose consideration may be
    well beyond human capabilities.
  • Furthermore, due to the subjectivity involved,
    the clustering procedure has to automatically
    derive the rules that assign cluster memberships
    to each data point. These cluster assignment
    rules have to be infered / learned from the
    available data.
  • Moreover, these rules have to learned without any
    feedback whether the rule is correct or
    incorrect. For this reason, clustering is
    considered an unsupervised learning task.

9
An Intro to Clustering
  • Clustering criteria are mainly based on data
    point similarity, such as intra-group and
    between-group point similarities. Many times
    relative inter-point distance between data points
    is used as a measure of similarity. Actually, in
    the examples of the 1st slide you unknowingly
    used relative inter-point distances to decide
    whether a collection of points forms a cluster or
    not!
  • When clusters are described by a representative
    data point (cluster prototype), such as the mean
    vector of all patterns belonging to the same
    cluster, clustering can be thought as a data
    compression effort the N original data points
    are replaced by the K cluster prototypes (KltltN).
    The collection of cluster prototypes is sometimes
    referred to as a codebook.
  • As a matter of fact, some image compression
    algorithms indeed use clustering at some stage of
    their operation.

10
The k-Means Clustering Algorithm
  • A very popular clustering procedure is the
    k-Means Algorithm.
  • It is a prototype-based algorithm using the mean
    vector (cluster center) of all data points
    belonging to a cluster as the cluster
    representative (thus its name).
  • A data point is assigned to a cluster if the
    cluster mean vector is the closest one among the
    other mean vectors (point to cluster points
    similarity).

11
The k-Means Clustering Algorithm
  • Heres roughly how it works
  • Initialization Randomly specify K cluster
    centers (Note each time you run k-Means you may
    get different clustering results!)
  • Until no data point changes cluster assignment do
    the following
  • Assign each point to the cluster whose center is
    closest (according to the usual Euclidean
    distance) to the point.
  • For each cluster re-compute its center as the
    mean vector of its data points.

12
The k-Means Clustering Algorithm
  • 2D Example
  • 31 data points
  • Original data shown prior to clustering

13
The k-Means Clustering Algorithm
  • 2D Example
  • K-Means run with K2.
  • Run completed after 2 iterations.
  • Cluster centers are indicated by crosses.

14
The k-Means Clustering Algorithm
  • 2D Example
  • K-Means run with K3.
  • Run completed after 3 iterations.
  • Cluster centers are indicated by crosses.

15
The k-Means Clustering Algorithm
  • 2D Example
  • K-Means run a 2nd time with K3.
  • Run completed after 5 iterations.
  • Cluster centers are indicated by crosses.
  • Notice that this time we got slightly
    different results.

16
The k-Means Clustering Algorithm
  • 2D Example
  • K-Means run with K6.
  • Run completed after 5 iterations.
  • Cluster centers are indicated by crosses.
  • Also relatively meaningful results.
  • Compression ratio 15

17
The k-Means Clustering Algorithm
  • Initiate MATLAB live demo here

18
An Intro to Classification
  • While during clustering data points are assigned
    cluster memberships in an unsupervised way,
    during classification (a.k.a pattern recognition)
    we are given data and their corresponding group
    memberships (class labels). This makes
    classification a supervised learning task.
  • The goal of classification is to identify rules
    from the available data points and their class
    labels (cumulatively known as the training data)
    so that additional, unseen, future data (termed
    test data) will be assigned the correct class
    labels (classified).
  • In practice we want to derive rules that will
    lead to the minimum misclassification error,
    because we may not be able to avoid errors
    (mislabelings).
  • The set of classification rules is a called a
    classification model or, simply, a classifier.

19
An Intro to Classification
  • Classification is a vast field, much richer and
    many times trickier to accomplish than
    clustering. We will not provide further details
    at this point.
  • Below is a pictorial example of a 2-class problem
    known as the noisy circle in the square. The
    classification task is to derive decision rules
    that discriminate points falling outside a red
    circle (blue) from the ones falling within
    (yellow). Notice that the model on the left gave
    more complicated rules that the one on the left
    and also makes more mistakes.

20
The ID3 Classification Algorithm
  • Here are some features of ID3
  • It can process data that are of mixed mode
    nature data point coordinates (a.k.a.
    attributes) can be numerical and/or categorical
    (e.g. red, tall, etc.)
  • The classification rules it produces are of the
    form of IF-THEN rules.
  • Moreover, the number of derived rules is kept to
    a minimum, which is highly desirable.
  • The classification rules are represented by a
    tree structure. Thus, ID3 is called a decision
    tree algorithm.

21
The ID3 Classification Algorithm
  • Consider an over-simplified, 2-class medical
    diagnosis related example
  • We have collected data from a number of potential
    patients. In specific, we recorded whether they
    feature a Symptom 1 (S1) and Symptom 2 (S2) and
    whether they have a specific disease (D1) or not
    (D0) D plays the role of the class label.
  • Each persons record consists of a triplet of
    values (S1, S2, D). 0 (1) signifies absence
    (presence) of the corresponding symptom. E.g.
    For a person that has Symptom 1 but not 2 that
    has the disease the recorded measurement is
    (1,0,1).
  • The rules that need to be infered from the
    available data are of the form IF S1x AND S2y
    THEN Dz. For this simple problem there are only
    8 class assignment rules to be considered (all
    0-1 combinations of x,y,z).

22
The ID3 Classification Algorithm
  • Lets assume the available data imply the
    following, apparent decision tree
  • The number of IF-THEN rules equal the number of
    terminal nodes of the tree. Each downward route
    from the top (root) node towards a terminal node
    represents a classification rule.

23
The ID3 Classification Algorithm
  • ID3 will generate the following decision tree
  • ID3 will discover that attribute S1 is irrelevant
    for correct classification and creates only 2
    rules.
  • While this example was too simple and we have not
    explained how ID3 infers its rules, examining all
    possible decision trees for large datasets and
    many attributes is almost impossible for humans.
    In these cases, ID3 usually provides a reasonable
    solution that leads to small classification error.

24
Genetic Algorithms in ML
  • Genetic Algorithms (GAs) is a biologically-inspire
    d, stochastic heuristic for solving combinatorial
    problems (e.g. find the combination that gives
    you best results).
  • In clustering, GAs may aid in inferring the best
    (in some predefined sense) cluster assignments.
  • In classification, GAs may aid in identifying the
    appropriate combination of decision rules that
    minimizes the misclassification error for a
    particular problem.

25
Genetic Algorithms in ML
  • The idea behind GAs is to treat the combinatorial
    solution (the combination) like a DNA string of
    an individual.
  • For example, consider the solution to a
    clustering problem with 2 clusters and 4 data
    points p1, p2, p3, p4, that assigned to clusters
    1, 2, 1, 2 respectively. Then, this solution
    (phenotype, in GA terminology) would correspond
    to a DNA string, a.k.a. genotype, of 1 2 1 2.
  • The GA starts off with a collection (population)
    of individuals featuring random genotypes and
    roughly performs the following steps
  • The fitness of each individual is evaluated
  • Randomly, but according to their fitness,
    individuals are selected for mating (selection
    step).
  • Individuals mate (chromosomal crossover step)
    giving birth to offsprings.
  • Offspring are randomly mutated (mutation step)
    with a low probability
  • Offspring replace their parents and form the next
    generation.

26
Genetic Algorithms in ML
  • The previous process is called evolution of the
    population.
  • The measure of fitness for each clustering
    solution would make sense to depend on how
    compact the produced clusters are and how well
    they are separated.
  • Below is an example of how crossover would work.
    The selected individuals are paired up. Then a
    random crossover point is selected and 2 new
    individuals are produced.

27
Genetic Algorithms in ML
  • In the mutation step a position in the genotype
    (a gene) is randomly selected with very small
    probability and is randomly changed like below
  • The amazing fact about GAs is that given a large
    population and having it evolve for many
    generations will eventually produce high-quality
    solutions (in our clustering example, good
    cluster assignments). And all this by imitating
    nature!
About PowerShow.com