home home

TECHNOLOGICAL EDGE

Zehu's Adaptive Speaker Verification technology offers the following advantages:

  • More accurate voice description
  • Multi-layer mechanism for classification
  • Cleansing and channel mismatch removal (de-noising)
  • Accelerated performance suitable for large scale environments

More Accurate Voice Description
Feature extraction is the first stage of the voice/speech recognition process, and Zehu technology utilizes different mathematical algorithms, including principle component analysis, heteroscedastic linear discrimination analysis, independent factor analysis and wavelets for this purpose. Zehu algorithms enhance common voice feature sets, thereby improving the speaker verification performance. This is possible because the Zehu algorithms find a set of features that correspond to a discriminate problem (discrimination of one speaker from anyone else) better than existing algorithms such as MFCC, LPC and PLP. The end result is a set of voice features that meet speaker verification requirements.

Multi-layer Mechanism for Classification
Zehu has developed a multi-layer model for recognizing the differences between different speakers - typically characterized as goats (speakers who are exceptionally unsuccessful at being accepted), sheep (speakers whose voice patterns are easily accepted), lambs (speakers who are exceptionally vulnerable to impersonation) and wolves (speakers exceptionally successful at impersonation). The aim of this multi-layer model is to cluster speakers according to these characteristics, while still recognizing differences between different speakers in the same cluster.

Cleansing and Channel Mismatch Removal (De-Noising)
When using any speaker recognition system in a real world environment, the major obstacle is the mismatch between the conditions during speech recording and the conditions under which the data to be recognized is recorded. The Zehu technology includes a mechanism to solve this problem, based on the integration of a series of algorithms which cover statistical classification and modeling of additional noises, statistical modeling of non-Gaussian noises, channel mismatch classification, de-noising modeling and double speech detection, echo cancellation and statistical voice-activity-based detection, and band pass filtering and adaptive RASTA,

Support Large Scale Environments and Accelerated Performance
Zehu's technology distributes loads between multiple servers, thereby enabling the support of large scale environments. Furthermore, it utilizes multiple cores (CPUs) running on the same machine, dramatically increasing performance.

 
    HOME | COMPANY | TECHNOLOGY | PRODUCTS | PARTNERS | NEWS/EVENTS | CONTACT US | SITE MAP  COPYRIGHT © 2008 ZEHU TECHNOLOGIES. SITE DESIGN: ZATAR CREATIVE