Browse Prior Art Database

Speech Recognition Method

IP.com Disclosure Number: IPCOM000121946D
Original Publication Date: 1991-Oct-01
Included in the Prior Art Database: 2005-Apr-04
Document File: 1 page(s) / 38K

Publishing Venue

IBM

Related People

Nishimura, M: AUTHOR

Abstract

This article describes a fenonic Markov model that can represent time- varying spectral patterns while using less memory space than existing models.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 91% of the total text.

Speech Recognition Method

      This article describes a fenonic Markov model that can
represent time- varying spectral patterns while using less memory
space than existing models.

      Formerly, we proposed a fenonic Markov model-based speech
recognizer that evaluates first-derivatives of spectral time
sequences D(t) and an instantaneous spectrum S(t) in each state of
the model independently [*].  In that model, the probability of a
pair of input label sequences LS and LD for a fenonic Markov word
model M was:
where MS (i)(or MD (i)) is a fenone at state i in the fenonic Markov
word model, I=i1  ,i2  ,...,it  is a state sequence, and B is a
transition from state i to j.  Although the model was effective for
representing time-varying spectral patterns, it still required many
parameters for its transition probabilities.

      In order to reduce the parameter size, we propose a new
formulation.  Given a fenonic Markov word model M, the probability of
the above input label sequences is estimated by the following
formula:

      It seems as if this model consists of two conventional models.
However, the two models are not trained or decoded separately.  In
spite of the approximation of independence in the transition, the
recognition accuracy in an experiment was almost the same as that of
the previous model.

      By using the new model, a 1000-word speech recognizer on a
speech processing card that has only 64K bytes of memory was
developed.

  ...