PSTL’s Speaker Diarization Patrick Nguyen Jean-Claude Junqua

Panasonic Speech Technology Laboratory

(PSTL)

Plan • System description • SWAMP

System description • • • •

Same as last year Segmentation: GMM decoding Five classes (narrow/wide x gender + sil) Heuristically smoothed

• + morph smoothing for zero-collar

System description • Same as last year • BIC with full covariance Gaussians

High risk novel approaches • Speaker diarization: – – – – –

Closed form solutions Quick test/train cycle Few components: no interaction Relatively new task Self-contained

Frontend and Speaker recognition • Standard question: why use STT features? • No convincing alternative: why? • Unlike linguistic content, speaker characteristics do not show on spectrogram • Unknown distortions in the spectrogram contain the information => Purely mathematical signal processing

SWAMP • SWAMP=Sweeping metric parameterization

• Input: MFCC • Output: Ranked SWAMP features • Basic idea: speech is a time curve on a

manifold

Metaphors: surface and grid

Z 21.0

t=4 -2 10.7

0

0.4

X -2

t=1

Y 0

t=2

2

t=3

2

Why Isometric Features? • Simple algebra and Gaussian OK

SWAMP: isometric

MFCC

Sweeping metric • Defines the shape of the squares [Scroedinger, Levin]

• Intuition: fast is imprecise Defined locally around y

Sweeping metric(2) • Each square has its metric g

Idea: triangulation • Characterize all points by relative distance • Flatten out surface

“Geodesic lengths”

Classical Metric Multidimensional Scaling (CMDS) • Visualization technique used in psychology mi

SB

Boston NYC

Dallas

SB

0

2663

2521

1326

0

190

1551

0

1373

Boston NYC Dallas

0

Isometric coordinates

Distances and MDS • Summary: – Find local metrics – Find geolen distances by integration

– SVD the distance matrix (MDS) mi

SB

Boston

NYC

Dallas

SB

0

2663

2521

1326

0

190

1551

0

1373

Boston NYC Dallas

0

Implementation details • Vector quantization: curse of dimensionality • Voronoi regions: locally linear regions • Adjacency: join squares of the surface • Tunneling: accidentally go “under” surface • Floyd-Warshall: complete paths

Vector Quantization • Curse of dimensionality prevents from computing local metrics (space not compactly populated) • Alternative (e.g. Levin): RiemannChristoffel symbol of the second kind

• Large ODE

Vector Quantization • Our solution: locally linear regions • Simplifies the geolen computation as well • Problem: VQ distortion is initialized with the contravariant coordinates

Voronoi regions • Polyhedral geodesics are straight lines if local embeddings are linear • Problem: Dirichlet/Voronoi interface is quadratic • Newton/Raphson solution of the Lagrangian Voronoi interface

Floyd-Warshall (or Dijkstra) • From local geolens to complete graph • Local geolens: Completed geolen

Tunneling and Adjacency • Use time structure to avoid going under the manifold

Z 21.0

-2 10.7

0

0.4

X -2 Y 0

2 2

Properties • Unique under any invertible constant distortion • Robust to quasi-stationary noise • Isometric • Time-distortion can be recovered by extending the vector (WRONG?) • Expensive

Future work? • Probablistic measure? Fisher? • Non-Riemannian topology? – Information divergence [Amari] – Renyi cross entropy?

• Spectral energy? • ASR?

Are segmentations created equal? Seg/Cluster

Spk Error

WER

CU/CU CU/PSTL

63.04% 18.51%

17.85%

BBN/BBN BBN/PSTL

~27% 19.33%

17.91%

CLIPS/CLIPS CLIPS/PSTL

57.91% 30.24%

LIA/LIA LIA/PSTL

19.21% 16.14%

LIMSI/LIMSI LIMSI/PSTL

35.10% 19.03%

MITLL2/MIT MITLL2/PSTL

20.04% 18.59%

PSTL/PSTL

18.58%

18.46%(ad)

17.56%

17.94%

Zero Collar • Zero collar requires precise speech/non speech • Morph filter minimizes the risk (trim a little from both ends) • Apply 0.3s bridging rule • (old system): 19.33% spkr error => 18.71%

Results • • • • • •

Zero collar is sharp Length of shows affects performance Last year: 18.8% spkr time err This year: 13.5% (18.6% relative) Only participant (PSTL) last year Lowest error

Conclusion • Novel approach (SWAMP) – Sweeping metric – Geolen computation – Multidimensional scaling

• Simple full covariance BIC clusterer

NIST / RT-2002 workshop PSTL

Scaling (CMDS). • Visualization technique used in psychology mi. SB. Boston NYC ... From local geolens to complete graph. • Local geolens: Completed geolen ...

464KB Sizes 1 Downloads 256 Views

Recommend Documents

NIST Academic Profile.pdf
Results 24 - 31 - Sweden. France. Israel. NZ. Taiwan. Others. 2017. ACT ACT Score. Section Summary. Middle 50%. Composite English Math Reading Science.

NIST SP800-68
Nov 2, 2005 - Network), Peter Tracy (Belarc), the Department of Energy, the Internal Revenue Service, and the Social. Security Administration. Additionally ...

NIST Giving Form.pdf
Wire transfer. Siam Commercial Bank PCL - Soi Chaiyot Branch ... NIST Giving Form.pdf. NIST Giving Form.pdf. Open. Extract. Open with. Sign In. Main menu.

NIST Student Medical Information.pdf
There was a problem previewing this document. Retrying... Download. Connect more ... NIST Student Medical Information.pdf. NIST Student Medical Information.

NIST Academic Profile.pdf
Results 22 - 32 - accreditation through the Council of International Schools (CIS) and New England Association of Schools and Colleges. (NEASC). Offering a ...

NIST SP800-68
Nov 2, 2005 - lists print and online resources that may be useful Windows XP security references. ... Windows XP security should take into account the role that the system plays. .... resources, and enable a password-protected screen saver. ...... Th

NIST 2017-2018 School Calendar.pdf
There was a problem previewing this document. Retrying... Download. Connect more apps... Try one of the apps below to open or edit this item. NIST 2017-2018 ...

NIST 2017 Annual General Report.pdf
NIST. 2017 ANNUAL GENERAL REPORT. GOVERNANCE. Message from the Chair of the NIST Foundation. Message from the Chair of the NIST Executive Board. NIST Foundation Board. Executive School Board. School Board Committees. The Role of the NIST Foundation.

Watch Sahel Door Nist (1962) Full Movie Online Free ...
Watch Sahel Door Nist (1962) Full Movie Online Free .Mp4______________.pdf. Watch Sahel Door Nist (1962) Full Movie Online Free .Mp4______________.

Workshop Description
Development, Experimentation, and Testing of Innovative Spectrum Sharing ... environment is critical to validating spectrum sharing technology under realistic ...

CFD Simulations on Extinction of Co-Flow Diffusion Flames - NIST
energy dissipation rate (m2/s3) ... and alternative suppression systems. .... than the heat of combustion for that mixture (i.e. not enough energy is released to ..... W.L., “A research agenda for the next generation of performance-based design.

pdf-146\nist-handbook-150-20-2005-edition-national-voluntary ...
... apps below to open or edit this item. pdf-146\nist-handbook-150-20-2005-edition-national-vo ... nformational-technology-security-testing-common-c.pdf.

NIST guidelines on firewalls and firewall policy.pdf
Page 2 of 48. Guidelines on Firewalls and Firewall. Policy. Recommendations of the National. Institute of Standards and Technology. Karen Scarfone.

DRAFT NIST Special Publication 800-63C - Kantara Initiative
May 18, 2016 - assertions; authentication; credential service provider; digital authentication; ..... checks the digital signature to verify that it was issued by a.

DRAFT NIST Special Publication 800-63C - Kantara Initiative
May 18, 2016 - entities, materials, or equipment are necessarily the best available for the purpose. ... The Information Technology Laboratory (ITL) at the National Institute ... mentioning or excluding others, or that a certain course of action is.

NIST Big Data Interoperability Framework: Volume 8 ... - GitHub
security and privacy of other than national security-related information in federal ...... introduce a number of objects that build the core of the interface addressing a .... While in physics the protection of the data is less of an issue, it is s s

A Review Study of NIST Statistical Test Suite
Development of an indigenous Computer Package .... A concept of degrees of freedom is introduced in these tests in the form of blocks or classes. For such ...

NIST Summer School Bus Transport Request.pdf
There was a problem previewing this document. Retrying... Download. Connect more apps... Try one of the apps below to open or edit this item. NIST Summer ...

Workshop chairs Workshop aim: gain insight on ...
Sociable Media to Support Social Connectedness. Workshop chairs. Thomas ... the types of content mediated by social media for this goal ... design experiences ...