×

Hidden markov model for speech processing with training method

  • US 9,020,816 B2
  • Filed: 08/13/2009
  • Issued: 04/28/2015
  • Est. Priority Date: 08/14/2008
  • Status: Active Grant
First Claim
Patent Images

1. A computerized method of detecting non-language speech sounds in an audio signal, comprising:

  • realizing with a computer a hidden Markov model comprising a plurality of states,wherein at least one of the plurality of states is associated with a non-language speech sound;

    isolating a segment of the audio signal;

    extracting a first feature set consisting of mel-frequency cepstral coefficients (MFCCs), pitch confidence, cepstral stationarity, and cepstral variance from the segment;

    using the first feature set to associate the segment with one or more of the plurality of states of the hidden Markov model; and

    classifying the segment as a language speech sound or a non-language speech sound accordingly.

View all claims
  • 1 Assignment
Timeline View
Assignment View
    ×
    ×