Method for word spotting in continuous speech

• US 5,425,129 A
• Filed: 10/29/1992
• Issued: 06/13/1995
• Est. Priority Date: 10/29/1992
• Status: Expired due to Fees

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Garman, Joseph H., Stanford, Vincent M., Klein, Alice G.
• Primary Examiner(s): Knepper, David D.
• Application Number:
  US07/968,097
• Time in Patent Office: 957 Days
• Field of Search: 381/43, 381/41-43, 395/2, 395/2.6-2.66, 395/2.56-2.59
• US Class Current: 704/256
• CPC Class Codes: G10L 15/142 Hidden Markov Models [HMMs]

Abstract

A digitized speech data channel is analyzes for the presence of words or phrases from a desired list. If the word is not present the speech time series from the channel, there is a high probability of matching the input data to phonemic noise markers rather than to word models. This is accomplished through the use of a competitive algorithm wherein Hidden Markov Models (HMMs) of the desired words or phrases compete with an alternative HMM of a set of phonemes. The set of phonemes can be chosen in order to reduce the computer resources required for the channel analysis.

Claims

1. In a speech recognition system, a subsystem for spotting pre-specified words, comprising:

a sub-word model analyzer using context independent phoneme models and having an input connected to a source of continuous speech;

a full-word model analyzer using triphone based models which may coincide with the context independent phoneme models, the full-word model analyzer [and]having an input connected to said source of continuous speech;

said sub-word model analyzer having a first threshold output signal indicative of the presence of one or more said phoneme models;

said full-word model analyzer having a second threshold indicative of the probability of the presence of a pre-specified word, represented by a triphone based model; and

a pre-specified word detecting means having an input coupled to said sub-word model analyzer, for receiving said first threshold output, and coupled to said full-word model analyzer, for receiving said second threshold, and in response thereto, identifying a pre-specified word to be spotted, provided the probability of the presence of the pre-specified word, represented as a triphone model is greater than the probability of the presence of the phoneme model.

2. The system of claim 1 wherein a reduced set of phoneme models is used to achieve improved run time efficiency for the system.

3. The system of claim 2 wherein the reduced set of phoneme comprises:

/S/,/T/,/N/,/UW/,/IY/,/AA/, and /!/.

4. The system of claim 3 wherein the identification of a prespecified word activates a voice-operated computer, which in the absence of such identification, remains in a quiescent state.

5. In a speech recognition system, a method for spotting pre-specified words, comprising the steps of:

analyzing with a sub-word model analyzer using context independent phoneme models, a source of continuous speech;

analyzing with a full-word model analyzer using triphone based models which may coincide with the context independent phoneme models, said source of continuous speech;

providing from said sub-word model analyzer a first threshold output signal indicative of the presence of one or more said phoneme models;

providing from said full-word model analyzer a second threshold output signal indicative of the probability of the presence of a pre-specified word, represented by a triphone based model;

detecting a pre-specified word with a pre-specified word detecting means, having an input coupled to said sub-word model analyzer, for receiving said first threshold signal, and coupled to said full-word model analyzer, for receiving said second threshold signal; and

identifying a pre-specified word in said source of continuous speech, provided the probability of the presence of the prespecified word is greater than the probability of the presence of the phoneme model.

6. The method of claim 5 wherein the step of analyzing with a sub-word analyzer uses a reduced set of phoneme models to achieve improved run time efficiency for the method.

7. The method of claim 6 wherein the reduced set of phonemes comprises:

/S/,/T/,/N/,/UW/,/IY/,/AA/, and /!/.

8. The method of claim 7 wherein the step of identifying a pre-specified word activates a voice-operated computer, which in the absence of such identification, remains in a quiescent state.