On the Use of PLDA i-Vector Scoring for Clustering Short Segments - - PowerPoint PPT Presentation

▶

Jan 16, 2024 347 likes •610 views

On the Use of PLDA i-Vector Scoring for Clustering Short Segments Itay Salmun Irit Opher Itshak Lapidot itshakl@afeka.ac.il itaysa@afeka.ac.il irito@afeka.ac.il Outline Shortly about DNNs Motivation Problem Definition Basic

SLIDE 1

On the Use of PLDA i-Vector Scoring for Clustering Short Segments

Itay Salmun Irit Opher Itshak Lapidot

itaysa@afeka.ac.il irito@afeka.ac.il itshakl@afeka.ac.il

SLIDE 2

Outline

Shortly about DNNs
Motivation
Problem Definition
Basic Mean-Shift Algorithm
Modified Mean-Shift Algorithm
Speaker Clustering System
Experiments and Results
Summary

June 2016

SLIDE 3

Shortly about DNNs

June 2016

SLIDE 4

Motivation

June 2016

Bli… Blu… Blo… Pla… Pli… Dla… Gla… Tra… Tra Ta Ta Ta Ta…

SLIDE 5

Motivation (cont.)

June 2016

Bli… Blu… Blo… Pla… Pli… Dla… Gla… Tra… Tra Ta Ta Ta Ta…

SLIDE 6

Motivation (cont.)

June 2016

Bli… Blu… Blo… Pla… Pli… Dla… Gla… Tra… Tra Ta Ta Ta Ta…

SLIDE 7

Problem Definition

June 2016

Given many short speech segments,

required to cluster them into homogeneous groups, such that:

– Each cluster will occupied mostly by one speaker only (cluster purity). – Each speaker will mostly belongs to one cluster

nly (speaker purity).

SLIDE 8

June 2016

Mean-Shift Algorithm

Basic

Objective : Find the densest

region

The Mean Shift vector:

Region of interest Center of mass Mean Shift vector

2 1 2 1

( )

n i i i h n i i

g h m g h φ φ φ φ φ φ φ

= =

  −     = −   −    

∑ ∑

The uniform kernel with bandwidth for Euclidean pairwise distances :

2 2 2 2

1 ( , , )

i i i

h g h h φ φ φ φ φ φ  − ≤  =  − >  

SLIDE 9

June 2016

Mean-Shift Algorithm (cont.)

Modified

The Mean Shift vector:
The adaptive bandwidth parameter is

calculated using K-Nearest neighbor

algorithm. If is the Kth nearest neighbor of

then the bandwidth is calculated as:

Where is the two-covariance scoring.

1 1

( , , ) ( ) ( , , )

k il il i i l h i i k il i i l

g h m g h φ φ φ φ φ φ φ

= =

= −

∑ ∑

( , )

i i iK

h s φ φ =

φ ( , )

i ik

s φ φ

SLIDE 10

June 2016

Mean-Shift Algorithm (cont.)

Modified

We select a subset of data points in

which the PLDA pairwise score with are larger or equal to the adaptive bandwidth

We use Mean shift weighted kernel of:

( ) { : ( , ) }

h i il i il i

S s h φ φ φ φ = ≥ ( , ) ( , ) ( , , ) ( , )

i il i il i i il i i il i

s s h g h s h φ φ φ φ φ φ φ φ ≥  =  <  ( )

h i

S φ

1 2 1 2 1 2

( , | ) ( , ) log ( , | )

s d

p H s p H φ φ φ φ φ φ =

SLIDE 11

June 2016

Speaker Clustering System

“i-vectors” PLDA score Mean Shift algorithm Clustering results

* In previous work: I. fixed h threshold II. a cosine distance instead of PLDA

III. Random Mean Shift

SLIDE 12

June 2016

Speaker Clustering System

Before clustering:

Train the UBM and TV matrix.
Train the PCA matrix T and the Whitening

transformation matrix C.

Calculate the low rank i-vectors:
Using the low rank i-vectors, train the two-

covariance model parameters. CT CT φ ϕ φ =

SLIDE 13

June 2016

Speaker Clustering System

Given a set of speech segments, cluster them according to the following steps:

1. For each speech segment extract the i-

vectors:

2. Calculate low rank i-vectors:
3. Apply two-covariance score mean-shift.
4. Merge all shifted points, according to

Euclidian distance with fixed threshold

{ }

SLIDE 14

June 2016

Experiments and Results

Experiments Setup

Experiments on telephone conversations

Cutting NIST-2008 into segments according to a given statistic.
Average segment length: 2.5 Sec
Average number of segments per speaker: 33

Clustering evaluation

1. Average Speaker Purity (ASP).
2. Average Cluster purity (ACP).
3. K:

4. Average Number of Detected Speakers (ANDS).

K acp asp = ⋅

SLIDE 15

June 2016