Patent · US Active

Methods, systems, and circuits for speaker dependent voice recognition with a single lexicon

US9633652B2 · kind B2 · utility

44Cited by
2References
13Claims
0Family size

Assignee

Inventors

Key dates

Filing dateMar 31, 2013
Grant dateApr 25, 2017
Priority date
Expiry dateMar 31, 2033

Classification

  • Technology area (CPC G)Physics
  • CPC primaryG10L17/04
  • WIPO fieldComputer technology
  • WIPO sectorElectrical engineering

Abstract

Embodiments reduce the complexity of speaker dependent speech recognition systems and methods by representing the code phrase (i.e., the word or words to be recognized) using a single Gaussian Mixture Model (GMM) which is adapted from a Universal Background Model (UBM). Only the parameters of the GMM need to be stored. Further reduction in computation is achieved by only checking the GMM component that is relevant to the keyword template. In this scheme, keyword template is represented by a sequence of the index of best performing component of the GMM of the keyword model. Only one template is saved by combining the registration template using Longest Common Sequence algorithm. The quality of the word model is continuously updated by performing expectation maximization iteration using the test word which is accepted as keyword model.

Source: USPTO / EPO open patent data. Objective bibliographic and citation counts.