The IBM 2016 Speaker Recognition System Seyed Omid Sadjadi, Sriram - - PowerPoint PPT Presentation

IBM Research

The IBM 2016 Speaker Recognition System

Seyed Omid Sadjadi, Sriram Ganapathy, Jason Pelecanos

IBM Research

2

Outline

• Introduction
• Speaker Recognition System
• Experimental Setup
• Results
• Conclusions

IBM Research

3

Introduction

IBM Research

4

Recent Progress

• Major advancements over the past several years.
• SOTA i-vector systems use UBMs to estimate stats
• Previous Work:
• Gaussian mixture model UBM [Reynolds 1997]
• Phonetically-inspired UBM (PI-UBM) [Omar 2010]
• DNN-based phonetically-aware UBM [Lei 2014]
• TDNN-based UBM (full-covariance) [Snyder 2015]
• DNN bottleneck based features are also used in SOTA systems

[Heck 1998; Richardson 2015; Matějka 2016 ]

IBM Research

5

Objectives

• To share state-of-the-art results on the NIST 2010 SRE
• To present the key system components that helped us

achieve these results

• Speaker- and channel-adapted fMLLR based features
• A DNN acoustic model with a large number of senones (10k)
• A nearest-neighbor discriminant analysis (NDA) technique
• To quantify the contribution of each component

IBM Research

6

Speaker Recognition System

IBM Research

7

Speaker Recognition System

• Our i-vector based speaker recognition system
• Speaker- and channel-normalized fMLLR based features
• i-vectors are estimated using DNN senone posteriors (~10k)
• LDANDA based intersession variability compensation

i-vector Extraction Acoustic Feats. Speech SAD Suff. Stats Dim. Reduc. T matrix LDA/NDA Score fMLLR PLDA

IBM Research

8

Feature space MLLR

Speech i-vector Extraction Acoustic Feats. SAD Suff. Stats Dim. Reduc. T matrix LDA/NDA Score fMLLR PLDA

IBM Research

9

DNN Senone I-vectors

Posteriors Senones (10k) B-W Statistics

i-vector Extraction Acoustic Feats. Speech SAD Dim. Reduc. T matrix LDA/NDA Score fMLLR PLDA Suff. Stats

IBM Research

• LDA assumes unimodal and Gaussian distributions
• It cannot effectively handle multimodal data
• It can be rank deficient

10

Linear Discriminant Analysis (LDA)

i-vector Extraction Acoustic Feats. Speech SAD Suff. Stats Dim. Reduc. T matrix LDA/NDA Score fMLLR PLDA

IBM Research Class 1 Class 2

Nearest Neighbor Discriminant Analysis (NDA)

global class means local k-NN means LDA NDA emphasize samples near boundary



( )( )

1 1 1

i

N C C T ij i ij i ij b l l l l l i j l j i

w

= = = ≠

= − −

∑∑∑

S x x M M

( ) ( )

{ }

( ) ( )

min , ( , ) , , ( , ) , ( , ) , ( , )

i i i i l K l l K l ij l i i i i l K l l K l

d NN i d NN j w d NN i d NN j

α α α α

= + x x x x x x x x

( )( )

1 C T b i i i i

p

=

= − −

∑

Sμ μ μ μ

11

IBM Research

12

Experimental Setup

IBM Research

13

Data

• Training Data
• NIST 2004-2008 SRE (English telephony and microphone data)
• Switchboard (SWB) cellular Parts I and II, SWB2 Phases II and III
• Total of 60,178 recordings
• Evaluation data
• NIST 2010 SRE (extended evaluation set)

Cond. Enroll Test Mismatch #Targets #Impostors C1

• Int. mic.
• Int. mic. (same type)

No 4,034 795,995 C2

• Int. mic.
• Int. mic. (different type)

Yes 15,084 2,789,534 C3

• Int. mic.

Telephony Yes 3,989 637,850 C4

• Int. mic.

Room microphone Yes 3,637 756,775 C5 Telephony Telephony (different type) Yes 7,169 408,950

IBM Research

14

DNN System Configuration

• 6 fully connected hidden layers with 2048 units
• The bottleneck layer has 512 units
• Trained using 600 hours of speech from Fisher
• Input is a 9-frame context of 40-D fMLLR feats.
• Estimates posterior probabilities of 10k senones
• 2k and 4k posteriors are also explored

IBM Research

15

Speaker Recognition System Configuration

• 500-dimensional total variability subspace trained using a subset
• f 48,325 recordings from NIST SRE, SWBCELL, and SWB2
• Sufficient statistics are generated using posteriors from:
• Gender independent 2048-component GMM-UBM (21,207 recordings)
• DNN with 7 hidden layers and 2k, 4k, or 10k senones
• MFCCs and fMLLR based features are evaluated
• LDA/NDA is applied to obtain 250-dimensional feature vectors
• Gaussian PLDA backend trained with 60,178 speech segments
• Evaluation metrics: Equal error rate (EER) and minDCF’08,’10

IBM Research

16

Results

IBM Research

17

LDA vs NDA (MFCC, 2048-GMM, 10k DNN, C5)

• NDA outperforms LDA for both GMM and DNN based systems

System EER [%] minDCF08 minDCF10 GMM-MFCC-LDA 2.40 0.12 0.439 GMM-MFCC-NDA 1.55 0.076 0.286 DNN-MFCC-LDA 1.02 0.045 0.168 DNN-MFCC-NDA 0.76 0.036 0.147

IBM Research

18

• Speaker- and channel-normalized fMLLRs outperform MFCCs

System EER [%] minDCF08 minDCF10 DNN-MFCC-LDA 1.02 0.045 0.168 DNN-fMLLR-LDA 0.82 0.032 0.120 DNN-MFCC-NDA 0.76 0.036 0.147 DNN-fMLLR-NDA 0.67 0.028 0.092

MFCC vs fMLLR (10k DNN, C5)

IBM Research

19

Impact of #Senones (fMLLR, C5)

• Using 10k senones gives the best performance
• NDA consistently outperforms LDA for 2k, 4k, and 10k senones
• Note: in contrast to DNNs, increasing the number of

components in GMMs (beyond 2k, with diag. cov. matrices) does not improve the results [Lei 2014; Snyder 2015].

System #Senones EER [%] minDCF08 minDCF10 DNN-LDA 2k 1.19 0.054 0.212 DNN-NDA 0.95 0.043 0.166 DNN-LDA 4k 0.98 0.041 0.169 DNN-NDA 0.86 0.033 0.116 DNN-LDA 10k 0.82 0.032 0.120 DNN-NDA 0.67 0.028 0.092

IBM Research

20

DET Plot Performance (C5)

IBM Research

21

System Progression (C5)

System EER [%] minDCF08 minDCF10 GMM-MFCC-LDA 2.40 0.120 0.439 1.55 0.076 0.286 0.76 0.036 0.147 0.67 0.028 0.092

• Achieved the best published performance (EER = 0.67%) on

NIST 2010 SRE (C5).

• Building upon previous best results:
• EER = 1.09% [Snyder 2015]  Gender-dependent (both genders)
• EER = 0.94% [Matějka 2016]  Gender-dependent (female trials)

IBM Research

22

Conclusions

IBM Research

23

• Presented the IBM i-vector speaker recognition system:

 Speaker- and channel-normalized fMLLR based features may be more effective than raw MFCCs in matched conditions  Using a DNN-UBM with 10k senones to partition the acoustic space provides the best performance  NDA more effective than LDA for channel compensation in the i-vector space (with multimodal data)

• Achieved the best published performance (EER = 0.67%)
• n NIST 2010 SRE (C5)
• For further progress on our system see us at IS-2016

Conclusions