Speech recognizer having a speech data memory storing speech data and a reference pattern memory storing partial symbol trains of words for recognition

• US 5,956,677 A
• Filed: 05/10/1996
• Issued: 09/21/1999
• Est. Priority Date: 05/12/1995
• Status: Expired due to Fees

• Current Assignee: NEC Corporation
• Original Assignee: NEC Corporation
• Inventors: Iso, Ken-ichi
• Primary Examiner(s): Hudspeth, David R.
• Assistant Examiner(s): Opsasnick, Michael N.
• Application Number:
  US08/644,273
• Time in Patent Office: 1,229 Days
• Field of Search: 395/2.63-2.67, 395/2.4, 395/2.45, 395/2.47, 704/254-258
• US Class Current: 704/254
• CPC Class Codes: G10L 15/06 Creation of reference templates; Training of speech recognition systems, e.g. adaptation to the characteristics of the speaker's voice G10L15/14 takes precedence
• CPC Class Codes: G10L 15/10 using distance or distortion measures between unknown speech and reference templates
• CPC Class Codes: G10L 15/142 Hidden Markov Models [HMMs]
• CPC Class Codes: G10L 15/1815 Semantic context, e.g. disambiguation of the recognition hypotheses based on word meaning
• CPC Class Codes: G10L 15/187 Phonemic context, e.g. pronunciation rules, phonotactical constraints or phoneme n-grams

Abstract

A speech recognizer includes a feature extracting unit for analyzing an input speech to extract a feature vector of the input speech. A speech data memory stores speech data and symbol trains of the input speech. A reference pattern memory stores sets each of a given partial symbol train of a word presented for recognition and an index of speech data with the expression thereof containing the partial symbol train in the speech data memory. A distance calculating unit reads out speech data corresponding to a partial symbol train stored in the reference pattern memory from the speech data memory, and calculates a distance between the partial symbol train read out from the reference pattern memory and a particular section of the input speech. A pattern matching unit derives, with respect to each word presented for recognition, a division of the subject word interval which minimizes the sum of distances of the input speech sections over an entire word interval. A recognition result calculating unit outputs, as a recognition result, a word presented for recognition, which gives the minimum one of the distances between the input speech data output of the pattern matching unit and all the words presented for recognition.

Claims

1. A speech recognizer comprising:

a speech data memory in which unfragmented speech data and symbol trains thereof are stored;

a reference pattern memory in which are stored sets, each of a given partial symbol train of a word presented for recognition, and an index of speech data with an expression thereof containing the given partial symbol train in the speech data memory, wherein no speech data is stored in the reference pattern memory;

a feature extracting unit for receiving an input speech signal, for extracting spectral features from the input speech signal, and for outputting the extracted spectral features as a spectral feature train;

a distance calculating unit for calculating a distance between the given partial symbol train stored in the reference pattern memory and the spectral feature train received from the feature extracting unit; and

a pattern matching unit for selecting, among all possible combinations of partial symbol trains as divisions of the symbol train of a word presented for recognition, a partial symbol train which minimizes a sum of distances of input speech sections over an entire input speech interval of the input speech signal, and outputting the sum of distances as data representing a distance between the given input speech section and the word presented for recognition,

wherein the selected partial symbol train is output as a word recognition result of the speech recognizer.

2. The speech recognizer according to claim 1, wherein the distance to be calculated in the distance calculating unit is the distance between the partial train of symbol train expression of speech data stored in the speech data memory and the given input speech section.

3. A speech recognizer comprising:

a feature extracting unit for analyzing an input speech to extract a feature vector of the input speech;

a speech data memory in which unfragmented speech data and symbol trains thereof are stored;

a reference pattern memory in which are stored sets, each of a given partial symbol train of a word presented for recognition, and an index of speech data with the expression thereof containing the partial symbol train in the speech data memory, wherein no speech data is stored in the reference pattern memory;

a distance calculating unit for reading out speech data corresponding to a partial symbol train stored in the reference pattern memory from the speech data memory and calculating a distance between the partial symbol train read out from the reference pattern memory and a particular section of the input speech;

a pattern matching unit for deriving, with respect to each word presented for recognition, a division of the subject word interval which minimizes the sum of distances of the input speech sections over an entire word interval; and

a recognition result calculating unit for outputting, as a recognition result, a word presented for recognition, which gives the minimum one of the distances between the input speech data output of the pattern matching unit and all the words presented for recognition.

4. A speech recognizing method, comprising the steps of:

a) storing, in a reference pattern memory, a plurality of reference patterns respectively corresponding to a plurality of words for recognition, each of said plurality of reference patterns corresponding to a particular sequence of phoneme symbols of a corresponding one of the plurality of words for recognition;

b) receiving an input speech signal and extracting a plurality of feature vectors as a feature vector train of the input speech signal;

c) partitioning each word for recognition into a plurality of word for recognition phone symbols, the plurality of word for recognition phone symbols for each word for recognition being partitioned into the particular sequence;

d) forming sets of said plurality of word for recognition phone symbols, each of said sets having said plurality of word for recognition phone symbols having the particular sequence as in the step c), one of said sets corresponding to said plurality of word for recognition phone symbols as partitioned in the step c), all others of said sets corresponding to at least one combination of two sequential ones of said plurality of word for recognition phone symbols into a combined word for recognition phone symbol;

e) creating a candidate word for recognition for each of said sets formed in the step d) by combining respective acoustical segments of reference speech data that is obtained from a speech data memory;

f) comparing said feature vector train to a plurality of candidate words for recognition respectively corresponding to different combinations of said plurality of phone symbols stored in said reference pattern memory to determine a corresponding sum of distance data, said sum of distance data representing a distance between said said feature vector train and said different combinations of said plurality of phone symbols stored in said reference pattern memory; and

f) determining which one of said different combinations of said plurality of phone symbols stored in said reference pattern memory has a minimum distance of said sum of distance data, said one being output as a most-likely candidate word for each word for recognition,

wherein the most-likely candidate word is a speech recognition output for the input speech signal.

5. The method according to claim 4, wherein each of the plurality of word for recognition phone symbols has a fixed length.

6. The method according to claim 5, wherein the combined word for recognition phone symbol has a length at least twice as long as the fixed length.

7. A speech recognizer, comprising:

an input unit configured to receive an input speech signal;

a feature extracting unit communicatively coupled to the input unit and configured to extract feature vectors from the input speech signal and to output a feature vector train as a result;

a speech data memory configured to store speech data;

a reference pattern memory configured to store sets of phoneme symbol trains as divisions of a plurality of words for recognition, wherein each of the sets corresponds to a particular sequential pattern of phoneme symbol trains combined in a different manner with respect to all other sets for the particular sequential pattern of phoneme symbol trains;

a distance calculating unit communicatively coupled to the reference pattern memory and the speech data memory, the distance calculating unit configured to retrieve all possible sets of phoneme symbol trains from the reference pattern memory for a word for recognition, and to obtain corresponding partial words from the speech data memory that make up each of the retrieved sets, and configured to create a reference word candidate based on the obtained partial words for each of the retrieved sets and to calculate a distance between the feature vector train of the input speech signal and each reference word candidate;

a pattern matching unit communicatively coupled to the distance calculating unit and the reference pattern memory, the pattern matching unit configured to determine which of the reference word candidates has a minimum distance with respect to the feature vector train of the input speech signal; and

a recognition result calculation unit communicatively coupled to the pattern matching unit and configured to output, as a word for recognition, one of the reference word candidates having the minimum distance with respect to the feature vector train of the input speech signal.