Dismiss
InnovationQ will be updated on Sunday, Oct. 22, from 10am ET - noon. You may experience brief service interruptions during that time.
Browse Prior Art Database

Silence/Mumble Model for Speech Recognition

IP.com Disclosure Number: IPCOM000112717D
Original Publication Date: 1994-Jun-01
Included in the Prior Art Database: 2005-Mar-27
Document File: 2 page(s) / 46K

Publishing Venue

IBM

Related People

Bahl, LR: AUTHOR [+5]

Abstract

Disclosed is an invention that improves the success rate of a voice command recognition system in noisy conditions, particularly when the noise is constituted of background speech (such as Public Address announcements). This is achieved by introducing a "silence/mumble" model as in the Figure.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 54% of the total text.

Silence/Mumble Model for Speech Recognition

      Disclosed is an invention that improves the success rate of a
voice command recognition system in noisy conditions, particularly
when the noise is constituted of background speech (such as Public
Address announcements).  This is achieved by introducing a
"silence/mumble" model as in the Figure.

                     ________
                       |        |
                   |->-| WORD_1 |->-|
      _________    |   |________|   |    _________
     |         |   |                |   |         |
 O->-| SIL/MUM |->-O                O->-| SIL/MUM |->-O
     |_________|   |    ________    |   |_________|
                   |   |        |   |
                   |->-| WORD_2 |->-|
                       |________|

      The Markov model SIL/MUM [1] will match both silence and
"mumble" (noise) better than any of the words in the vocabulary.

The "silence/mumble" model probabilities can be derived, for example,
as follows:

      The system is trained on clean training data, i.e., data that
does not contain any mumbles.  In particular, the training will
provide parameter values for a SILENCE model [2].

      Let p(s, l) denote the trained output probability for label l
in state s of the model for...