Title: Robust Voice Activity Detection for Interview Speech in NIST Speaker Recognition Evaluation
1Robust Voice Activity Detection for Interview
Speech in NIST Speaker Recognition Evaluation
- Man-Wai MAK and Hon-Bill YU
- The Hong Kong Polytechnic University
- enmwmak_at_polyu.edu.hk
- http//www.eie.polyu.edu.hk/mwmak/
2Outline
- Speaker Verification
- Speaker Verification Process
- Voice Activity Detection (VAD) in Speaker
Verification - Effect of VAD on Acoustic Features
- Characteristics of Interview-Speech in NIST
Speaker Recognition Evaluation - VAD for NIST Speaker Recognition Evaluation
- Experiments on NIST SRE 2008
- Preliminary Results on NIST SRE 2010
3Speaker Verification Process
- To verify the identify of a claimant based on
his/her own voices
Is this Marys voice?
I am Mary
4Speaker Verification Process
- A 2-class Hypothesis problem
- H0 MFCC sequence X(c) comes from to the true
speaker - H1 MFCC sequence X(c) comes from an impostor
- Verification score is a likelihood ratio
Speaker Model
Score
Feature extraction
Decision
-
Background Model
5Voice Activity Detection in Speaker Verification
VAD
Feature Extraction
Speech
Acoustic Features (MFCC)
Speech segments
MFCC
DCT
LogX(?)
6Effect of VAD on Acoustic Features
Non-speech region
Feature Extraction
Feature Extraction
VAD
Speech
7Outline
- Speaker Verification
- Speaker Verification Process
- Voice Activity Detection (VAD) in Speaker
Verification - Effect of VAD on Acoustic Features
- Characteristics of Interview-Speech in NIST
Speaker Recognition Evaluation - VAD for NIST Speaker Recognition Evaluation
- Experiments on NIST SRE 2008
- Preliminary Results on NIST SRE 2010
8Interview-Speech in NIST SRE
Interviewee
Desk
Interviewer
Interview Room
Source NIST SRE 2008 Workshop
9Interview-Speech in NIST SRE
- Far-field and desktop microphones were used for
collecting interview speech - Some interview-speech files are very noisy,
causing difficulty in differentiating speech
segments from non-speech segments
non-speech
speech
Frequency
Amplitude
Time
A typical interview-speech file in NIST SRE 2008
10Interview-Speech in NIST SRE
- Some files have very low SNR
Whole file
Amplitude
Amplitude
S speech h non-speech
S speech
Segmentation
Frequency
Time
10
11Interview-Speech in NIST SRE
- Some files contain spiky signals, causing wrong
VAD decision threshold
Spiky signal
Amplitude
Time
12Interview-Speech in NIST SRE
- Some files contain low-energy speech signal
superimposed on periodic background noise.
Amplitude
Non-speech detected as speech
Segmentation
Frequency
Time
13Outline
- Speaker Verification
- Speaker Verification Process
- Voice Activity Detection (VAD) in Speaker
Verification - Effect of VAD on Acoustic Features
- Characteristics of Interview-Speech in NIST
Speaker Recognition Evaluation - VAD for NIST Speaker Recognition Evaluation
- Experiments on NIST SRE 2008
- Preliminary Results on NIST SRE 2010
14VAD for NIST Speaker Recognition Evaluation
- Use speech enhancement as a pre-processing step
Speech Segment Info
Denoising (Spectral Subtraction)
Energy-based VAD
Noisy Speech
Denoised Speech
Spectral-Subtraction VAD (SVAD)
Feature Extraction
Scoring
Decision Making
Accept/Reject
MFCC
S
S
S
S
S
S
Speaker Model
Decision Threshold
Impostor Model
15VAD for NIST Speaker Recognition Evaluation
- Use speech enhancement as a pre-processing step
Signal Frequency Spectrum
Clean speech x(n,m) X(?,m)
Noisy speech y(n,m) Y(?,m)
Background speech b(n,m) B(?,m)
This values were set such that we remove as much
noise as possible.
16VAD for NIST Speaker Recognition Evaluation
- Without denoising
- With denoising
Amplitude
Time
Amplitude
Time
17VAD for NIST Speaker Recognition Evaluation
S speech h non-speech
18VAD for NIST Speaker Recognition Evaluation
With denoising
SS-VAD
S speech h non-speech
VAD in ETSI-AMR speech coder
19VAD for NIST Speaker Recognition Evaluation
- Speech-segment-length to speech-file-length ratio
of 3 VADs
total duration 10 secs .
total speech segment 3 secs.
speech-segment-length to speech-file-length ratio
3/10
20VAD for NIST Speaker Recognition Evaluation
- Speech-segment-length to speech-file-length ratio
of 3 VADs
VAD in ETSI AMR Coder
Ordinary Energy-based VAD
Spectral-Subtraction VAD
High frequency of occurrence, suggesting many
non-speech segments being mistakenly detected as
speech segments
21Outline
- Speaker Verification
- Speaker Verification Process
- Voice Activity Detection (VAD) in Speaker
Verification - Effect of VAD on Acoustic Features
- Characteristics of Interview-Speech in NIST
Speaker Recognition Evaluation - VAD for NIST Speaker Recognition Evaluation
- Experiments on NIST SRE 2008
- Preliminary Results on NIST SRE 2010
22Experiments on NIST SRE 2008
- Speaker Modeling GMM-SVM
- Score Normalization T-norm
23Results on NIST 2008 SRE
24Results on NIST 2008 SRE
Common Condition 1
VAD
ETSI AMR
SS-VAD
25Preliminary Results on NIST 2010
Common Condition 2 All trials involving
interview speech from different microphones
EER () Normalized minDCF
Energy-based VAD 11.72 0.99
SS-VAD 4.45 0.58
SMB 5.83 0.75
SS-SMB 4.62 0.60
NIST ASR Transcripts 8.58 0.85
ETSI-AMR 8.05 0.85
SMB Statistical-Model Based VAD Sohn, et al. A
statistical model-based voice activity
detection, IEEE Signal Processing Letters, 1999.
26Conclusions
- Noise reduction is of primary importance for VAD
under extremely low SNR - It is important to remove the sinusoidal
background found in NIST SRE sound files as this
kind of background signal could lead to many
false detection in energy-based VAD. - Using noise reduction as a pre-preprocessing step
leads to a VAD outperforms the VAD in ETSI-AMR
(Option 2).
27VAD for NIST Speaker Recognition Evaluation
- Threshold Determination and VAD Decision Logic
Sample-based
Windowing
Frame-based
Amplitude Ranking
28Results
29Experiments on NIST SRE 2008
(NIST05 06)
MAP Adaptation
Feature Extraction
MAP Adaptation
300 background speakers (NIST06)
(NIST08)
GMM-supervectors of 300 impostors
GMM-supervectors of target speakers
NAP
NAP
30Experiments on NIST SRE 2008