Speech recognition training method

• US 4,718,088 A
• Filed: 03/27/1984
• Issued: 01/05/1988
• Est. Priority Date: 03/27/1984
• Status: Expired due to Term

• Current Assignee: Voice Industries Corporation
• Original Assignee: Exxon Research and Engineering Company (Exxon Mobil Corporation)
• Inventors: Lee, Chin-Hui, Ganesan, Kalyan, Klovstad, John W., Baker, James K.
• Primary Examiner(s): Kemeny, E. S. Matt
• Application Number:
  US06/593,891
• Time in Patent Office: 1,379 Days
• Field of Search: 381/41-43, 364/513.5
• US Class Current: 704/241
• CPC Class Codes: G10L 15/02 Feature extraction for speech recognition; Selection of recognition unit
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 15/083 Recognition networks G10L15/142, G10L15/16 take precedence
• CPC Class Codes: G10L 15/12 using dynamic programming techniques, e.g. dynamic time warping [DTW]
• CPC Class Codes: G10L 2015/0638 Interactive procedures
• CPC Class Codes: G10L 25/27 characterised by the analysis technique
• CPC Class Codes: G10L 25/87 Detection of discrete points within a voice signal

Abstract

A speech recognition method and apparatus employ a speech processing circuitry for repetitively deriving from a speech input, at a frame repetition rate, a plurality of acoustic parameters. The acoustic parameters represent the speech input signal for a frame time. A plurality of template matching and cost processing circuitries are connected to a system bus, along with the speech processing circuitry, for determining, or identifying, the speech units in the input speech, by comparing the acoustic parameters with stored template patterns. The apparatus can be expanded by adding more template matching and cost processing circuitry to the bus thereby increasing the speech recognition capacity of the apparatus. Template pattern generation is advantageously aided by using a "joker" word to specify the time boundaries of utterances spoken in isolation, by finding the beginning and ending of an utterance surrounded by silence.

Claims

1. In a speech recognition apparatus wherein speech units are each characterized by a sequence of template patterns, and having

means for processing a speech input signal for repetitively deriving therefrom, at a frame repetition rate, a plurality of speech recognition acoustic parameters, and

means responsive to said acoustic parameters

for generating likelihood costs between said acoustic parameters and said speech template patterns, and

for processing said likelihood costs for determining the speech units in said speech input signal,

a method for generating said template patterns comprising the steps of

finding the beginning and end of an input speech unit surrounded by silence for which template patterns are to be generated, and

generating in accordance with a known procedure, template patterns representing said speech unit,

said finding step comprising

modelling silence as a template pattern,

for each frame, comparing said silence template pattern likelihood cost with a fixed reference threshold value, and

declaring the beginning of said speech unit when the score for the silence template pattern crosses the threshold value.

2. The speech recognition template pattern generating method of claim 1 further comprising the step of

declaring the end of said speech unit when the score for the silence template improves sufficiently to cross a second threshold value.

3. The speech recognition template pattern generating method of claim 2 wherein said second threshold value is less than said first threshold value.

6. The speech recognition template pattern generating method of claim 1 further comprising the step of

associating with an intermediate node to which said first arc leads, a second arc having a second fixed reference threshold value, and

associating with a second intermediate node to which said second arc leads, a third arc corresponding to the likelihood score of silence, and

determining when said score at a node to which said third arc leads matches a predetermined condition.

7. The speech recognition template pattern generating method of claim 6 wherein said second threshold value is less than said first threshold value.

4. In a speech recognition apparatus wherein speech units are each characterized by a sequence of template patterns, and having

means for processing a speech input signal for repetitively deriving therefrom, at a frame repetition rate, a plurality of speech recognition acoustic parameters, and

means responsive to said acoustic parameters

for generating likelihood costs between said acoustic parameters and said speech template patterns, and

for processing said likelihood costs for determining the speech units in said speech input signal,

a method for generating said template patterns comprising the steps of

finding the beginning and end of an input speech unit surrounded by silence for which template patterns are to be generated, and

generating in accordance with a known procedure, template patterns representing said speech unit,

said finding step comprising

modelling silence as a template pattern,

for each frame, comparing said silence template pattern likelihood cost with a fixed reference threshold value, and

declaring the beginning of said speech unit when the score for the silence template pattern crosses the threshold value,

declaring the end of said speech unit when the score for the silence template improves sufficiently to cross a second threshold value, and

wherein said second threshold value is less than said first threshold value.

5. In a speech recognition apparatus wherein speech units are each characterized by a sequence of template patterns, and having

means for processing a speech input signal for repetitively deriving therefrom, at a frame repetition rate, a plurality of speech recognition acoustic parameters, and

means responsive to said acoustic parameters

for generating likelihood costs between said acoustic parameters and said speech template patterns, and

for processing said likelihood costs for determining the speech units in said speech input signal,

a method for generating said template patterns comprising the steps of

finding, using a dynamic programming and a grammar graph, the beginning and end of an input speech unit surrounded by silence for which template patterns are to be generated, and

generating in accordance with a known procedure, template patterns representing said speech unit,

said finding step comprising

modelling silence as a template pattern,

associating from a beginning node of said grammar graph a first arc having a fixed reference threshold value, and

associating with said beginning node a silence self loop, and following said dynamic programming for determining the beginning of said speech unit.