Dual-step sound pattern matching

• US 4,776,017 A
• Filed: 04/30/1986
• Issued: 10/04/1988
• Est. Priority Date: 05/01/1985
• Status: Expired due to Fees

• Current Assignee: Ricoh Company Limited
• Original Assignee: Ricoh Company Limited
• Inventors: Fujimoto, Junichiroh
• Primary Examiner(s): KEMENY, EMANUEL
• Application Number:
  US06/857,838
• Time in Patent Office: 888 Days
• Field of Search: 381/42, 381/43, 364/513.5
• US Class Current: 704/244
• CPC Class Codes: G10L 15/00 Speech recognition G10L17/00 takes precedence

Abstract

A sound or speech recognition system is made faster by first using data-reduced patterns to find candidates in preliminary testing, and then making a final decision by testing against full-data patterns. Specifically, a signal is transformed to a time-sequence of spectral vectors, which forms a full-data pattern. Then, a predetermined number of vectors representing the beginning or front-end of the signal are duplicated and summed to form a sum spectral vector which is used as the data-reduced pattern for preliminary testing.

Claims

1. A method for identifying an unknown sound having a registration mode and an identification mode, wherein, during the registration mode, said method comprising the steps of:

forming a time-dependent first pattern from a sound inputted to be registered; and

storing said first pattern as a first reference pattern while adding a predetermined portion of said first pattern together to define a second reference pattern and storing said second reference pattern, whereby first and second reference patterns are stored for various sounds to be registered;

during the identification mode, said method comprising the steps of;

forming a time-dependent second pattern from an unknown sound inputted to be identified;

adding a predetermined portion of said second pattern together to define a comparison pattern;

comparing said comparison pattern with said stored second reference patterns, thereby selecting those second reference patterns which have at least a predetermined level of similarity with said comparison pattern; and

3. The method of claim 1 wherein said time-dependent pattern is a time-frequency pattern which is formed by sampling sound data at a predetermined time interval at a plurality of different frequency intervals, which defines a frame, at the same time.

5. The method of claim 1 wherein said second reference pattern is normalized in a predetermined manner before being stored.

2. comparing said second pattern with those of said first reference patterns which correspond to the selected second reference patterns to determine which has the best similarity, thereby identifying the unknown sound.

4. The method of claim 2 wherein each of the steps of adding a predetermined portion adds a predetermined number of first frames in each of the time-dependent patterns.

6. The method of claim 4 wherein said comparison pattern is normalized in a predetermined manner before being used at the comparing step.

7. The method of claim 4 wherein said predetermined manner is a two-bit normalization.

9. The method of claim 7 wherein said time-dependent pattern is a time-frequency pattern which is formed by sampling sound data at a predetermined time interval at a plurality of different frequency intervals, which defines a frame, at the same time.

8. A method for identifying an unknown sound having a registration mode and an identification mode, wherein, during the registration mode, said method comprising the steps of:

forming a time-dependent first pattern from a sound inputted to be registered; and

storing said first pattern as a first reference pattern while adding a predetermined portion of said first pattern together to define a second reference pattern, classifying said second reference pattern into one of predetermined types, and storing the type of said second reference pattern, whereby first reference patterns and types thereof are stored for various sounds to be registered;

during the identification mode, said method comprising the steps of;

forming a time-dependent second pattern from an unknown sound inputted to be identified;

adding a predetermined portion of said second pattern together to define a third pattern;

classifying said third pattern into one of said predetermined types; and

comparing said second pattern with those of said first reference patterns which have the same type as that of said third pattern to determine which has the best similarity, thereby identifying the unknown sound.

10. The method of claim 8 wherein each of the steps of adding a predetermined portion adds a predetermined number of first frames in each of the time-dependent patterns.