Speech classifier and method using delay elements

• US 6,038,535 A
• Filed: 03/23/1998
• Issued: 03/14/2000
• Est. Priority Date: 03/23/1998
• Status: Expired due to Term

• Current Assignee: Google Technology Holdings LLC (Alphabet Inc.)
• Original Assignee: Motorola, Inc. (Motorola Solutions, Inc.)
• Inventors: Campbell, William Michael
• Primary Examiner(s): Dorvil, Richemond
• Application Number:
  US09/045,917
• Time in Patent Office: 722 Days
• Field of Search: 704/222, 704/232, 704/270, 704/275, 704/231, 704/236, 704/250, 704/246, 704/251, 704/255, 704/245, 704/230
• US Class Current: 704/275
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 15/08 Speech classification or search

Abstract

Classifiers (110) and a selector (112) perform an identification method (300) to identify an ordered set of vectors (e.g., spoken commands, phoneme identification, radio signatures, communication channels, etc.) representing a class as one of a predetermined set of classes. Training processor (104) performs a training method (200) to train a set of models and store the models in a model memory (108). Classifiers (110) receive models from the model memory and combine the models with the ordered set of vectors to determine a set of scores. The selector associates the set of scores with the predetermined set of classes to identify the ordered set of vectors as a class from the predetermined set of classes.

Claims

1. A speech classifier comprising:

a feature memory for storing feature vectors that represent predetermined verbal commands;

a training processor for generating command models from the feature vectors;

a command model memory for storing the command models;

a first and second plurality of time-delay elements for sequentially propagating each element of a set of received feature vectors, the set of received feature vectors representing an unidentified spoken command;

a first classifier element coupled with the first plurality of delay elements, the first classifier element combining the sequentially propagated feature vectors from the first plurality of delay elements with a first of the command models and providing a first series of scalar values corresponding to the sequentially propagated feature vectors;

a second classifier element coupled with the second plurality of delay elements, the second classifier element combining the sequentially propagated feature vectors from the second plurality of delay elements with a second of the command models and providing a second series of scalar values corresponding to the sequentially propagated feature vectors; and

a selector (112) for selecting one of the command models corresponding to one of the verbal commands based on the first and second series of scalar values.

2. The speech classifier as claimed in claim 1 wherein the training processor generates a predetermined number of command models from the feature vectors, and the model memory stores the predetermined number of command models,

and wherein the speech classifier further comprises;

a predetermined number of pluralities of time-delay elements, the predetermined number including said first and second pluralities of time-delay elements; and

a predetermined number of classifier elements, the predetermined number including the first and second classifier elements,

wherein the predetermined number of classifier elements and the predetermined number of pluralities of time-delay elements correspond with the predetermined number of command models.

3. A method of speech classification comprising the steps of:

storing feature vectors that represent predetermined verbal commands;

generating command models from the feature vectors;

sequentially propagating each element of a set of received feature vectors through a first and second plurality of time-delay elements, the set of received feature vectors representing an unidentified spoken command;

combining in a first classifier element the sequentially propagated feature vectors from the first plurality of delay elements with a first of the command models to provide a first series of scalar values corresponding to the sequentially propagated feature vectors;

combining in a second classifier element the sequentially propagated feature vectors from the second plurality of delay elements with a second of the command models to provide a second series of scalar values corresponding to the sequentially propagated feature vectors; and

selecting one of the command models corresponding to one of the verbal commands based on the first and second series of scalar values.

4. The method as claimed in claim 3:

wherein the generating step comprises generating a predetermined number of said command models from the feature vectors

wherein the sequentially propagating step further comprises the step of sequentially propagating each element of the received feature vectors through a predetermined number of pluralities of time-delay elements;

wherein the combining step further comprises the step of combining the sequentially propagated feature vectors from the time delay elements in a corresponding classifier element of a predetermined number of classifier elements,

and wherein a predetermined number of pluralities of time-delay element include said first and second pluralities of time-delay elements,

said predetermined number of classifier elements include the first and second classifier elements, and

wherein the predetermined number of classifier elements and the predetermined number of pluralities of time-delay elements correspond with the predetermined number of command models.