Speech recognizer

• US 5,337,394 A
• Filed: 06/09/1992
• Issued: 08/09/1994
• Est. Priority Date: 06/09/1992
• Status: Expired due to Term

• Current Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Original Assignee: Kurzweil Applied Intelligence, Inc. (Intel Corporation)
• Inventors: Sejnoha, Vladimir
• Primary Examiner(s): MacDonald, Allen R.
• Assistant Examiner(s): KIM, RICHARD
• Application Number:
  US07/895,618
• Time in Patent Office: 791 Days
• Field of Search: 381/41, 381/43, 395/2.5
• US Class Current: 704/241
• CPC Class Codes: G10L 15/10 using distance or distortion measures between unknown speech and reference templates

Abstract

In the speech recognizer disclosed herein, alignment of an unknown speech sediment, represented by a finely gradiated sequence of frames, with a model sediment represented by a sequence of states is performed by first preparing respective coarse sequences representing the unknown and model segments thereby to define a coarse matrix representing possible alignments. The fine sequences correspondingly define a fine matrix. A best alignment of the coarse sequences is determined thereby to define a coarse path through the coarse matrix. The coarse path is overlaid on the fine matrix and a corridor is defined which includes fine matrix locations which lie within a preselected metric of the coarse path. Only transitions within the corridor are calculated in determining the fine alignment of the unknown speech segment with the model segment, thereby significantly reducing the number of computations required.

Claims

1. In a speech recognizer which operates to match an unknown speech segment comprising a fine sequence of frames with model segments represented by respective fine sequences of states, the respective fine sequences of the unknown segment an a model segment together defining a fine matrix;

a method of determining a good alignment of the unknown segment with a model segment, said method comprising;

preparing respective coarse sequences representing said unknown speech segment and said model segment thereby to obtain a respective coarse matrix;

determining a best alignment of said coarse sequences thereby determining a coarse path through said coarse matrix;

overlaying said coarse path on said fine matrix and determining which fine matrix locations lie within a preselected metric of said coarse path thereby defining a corridor of possible paths through said fine matrix; and

calculating only transitions within said corridor, determining an alignment of said unknown segment with said model segment.

2. The method according to claim 1 wherein said coarse sequences are obtained by subsampling the respective fine sequences, skipping a predetermined number between subsamples.

3. The method according to claim 2 wherein said coarse sequences include the first and last items from the respective fine sequences.

4. The method according to claim 1 wherein the best alignment of said coarse sequences is determined by Viterbi decoding.

5. The method according to claim 4 wherein the alignment of fine sequences is obtained by Viterbi decoding constrained to transitions within said corridor.

6. The method according to claim 1 wherein said predetermined metric selects as being within said corridor only fine matrix locations which are within a predetermined number of frames of said coarse path.

7. In a speech recognizer which operates to match an unknown speech segment comprising a fine sequence of multi-dimensional spectral frames with model segments also represented by a respective fine sequences of states, the respective fine sequences of the unknown segment and a model segment together defining a fine matrix which encompasses all possible alignments of the unknown segment fine sequence with the respective model segment fine sequence;

a method of determining a good alignment of the unknown segment fine sequence with a model segment fine sequence, said method comprising;

preparing respective coarse sequences representing said unknown speech segment and said model segment thereby to obtain a coarse matrix representing possible alignments of said unknown and model segments;

determining a best alignment of said coarse sequences thereby determining a coarse path through said coarse matrix;

overlaying said coarse path on said fine matrix and interpolating between the locations defining said coarse path thereby to determine a base path across said fine matrix;

determining which fine matrix locations lie within a preselected metric of said coarse path thereby defining a corridor comprising a limited number of possible paths through said fine matrix; and

calculating only transitions within said corridor, determining a best alignment of said unknown segment fine sequence with said model segment fine sequence.

8. The method as set forth in claim 7 wherein said coarse sequences are obtained by subsampling the respective fine sequences.

9. The method as set forth in claim 8 wherein said predetermined metric selects as being within said corridor only fine matrix locations which are within a preselected number of frames of said base path.

10. In a speech recognizer which operates to match an unknown speech segment comprising a fine sequence of frames with model segments represented by respective fine sequences of states, the respective fine sequences of the unknown segment and a model segment together defining a fine matrix which encompasses all possible alignments of the unknown segment with the respective model segment;

a method of determining a good alignment of the unknown segment with a model segment, said method comprising;

subsampling the respective fine sequences thereby to obtain respective coarse sequences representing said unknown speech segment and said model segment and thereby to define a coarse matrix representing possible alignments of said unknown and model segments;

determining a best alignment of said coarse sequences thereby determining a coarse path through said coarse matrix;

overlaying said coarse path on said fine matrix and interpolating between the locations defining said coarse path thereby to determine a base path across said fine matrix;

determining which fine matrix locations lie within a preselected number of locations of said base path thereby defining a corridor of possible paths through said fine matrix; and

calculating only transitions within said corridor, determining a best fine alignment of said unknown segment with said model segment.