Method and neural network for speech recognition using a correlogram as input

• US 5,721,807 A
• Filed: 01/21/1994
• Issued: 02/24/1998
• Est. Priority Date: 07/25/1991
• Status: Expired due to Fees

• Current Assignee: Siemens AG
• Original Assignee: Siemens AG
• Inventors: Tschirk, Wolfgang
• Primary Examiner(s): MacDonald, Allen R.
• Assistant Examiner(s): MATTSON, ROBERT
• Application Number:
  US08/185,800
• Time in Patent Office: 1,495 Days
• Field of Search: 395/2.11, 395/2.41, 395/2.68, 395/2.64, 395/2.6-2.63, 395/2.75
• US Class Current: 704/255
• CPC Class Codes: G10L 15/16 using artificial neural networks

Abstract

A method and device for recognizing individual words of spoken speech can be used to control technical processes. The method proposed by the invention is based on feature extraction which is particularly efficient in terms of computing capacity and recognition rate, plus subsequent classification of the individual words using a neural network.

Claims

1. A method for recognizing individual words of speech, which comprises:

converting speech during an expectation time period into an electrical speech signal;

ascertaining an instantaneous spectral amplitude distribution of the speech signal during time intervals defined by a duration of a phoneme and representing the instantaneous spectral amplitude distribution as a spectral vector Sⁱ (i=0, 1, . . . , m-1), wherein each element (Sⁱ₀, Sⁱ₁, . . . , Sⁱ_n-1) of the spectral vector Sⁱ represents an amplitude of a frequency band having a predetermined bandwidth, and n is an integer representing a number of divisions of a total detected frequency band into the frequency bands having the predetermined bandwidth;

forming a spectogram S from the spectral vectors Sⁱ in accordance with ##EQU2## deriving a correlogram K from the spectrogram S, wherein the correlogram K has coordinates j, h, k and each element K_j,h,k of the correlogram K is formed in accordance with ##EQU3## and classifying an individual spoken word with a word-typical characteristic pattern with the correlogram K.

2. The method according to claim 1, which comprises defining the time interval t_s between two successive spectral vectors Sⁱ and Sⁱ⁺¹ equal to 32 ms.

3. The method according to claim 1, which comprises selecting each of the indices j, h, k of the correlogram K according to

4. The method according to claim 1, which comprises classifying the spoken individual word with a neural network.

5. The method according to claim 4, which comprises

assigning each element of the correlogram K to each of a first number of neurons in an input plane of the neural network;

assigning each of the neurons of the input plane to each of a second number of neurons in an output plane of the neural network; and

indicating a defined recognized individual word with an output of a respective one of the neurons of the output plane.

6. The method according to claim 4, which comprises calculating in each of the neurons with a nonlinear transfer function ##EQU4##

7. The method according to claim 1, which further comprises initiating a dialing process in a telephone set as a function of the recognized individual word of a telephone number associated with the individual word.

8. An apparatus for recognizing spoken words, comprising:

a digital signal processor connected to a bus system, said bus system including data, address and control lines, and a program memory, a working memory, and an input/output unit each connected to said digital signal processor via said bus system;

said digital signal processor including means for converting speech spoken during an expectation time period into an electrical speech signal;

means for ascertaining an instantaneous spectral amplitude distribution of the speech signal during a time interval defined by a duration of a phoneme;

means for representing the instantaneous spectral amplitude distribution as a spectral vector, each element of the spectral vector representing an amplitude of a frequency band having a predetermined bandwidth; and

means for classifying an individual spoken word with a word-typical characteristic pattern derived from the spectral vector representing the instantaneous spectral amplitude distribution of the individual spoken word.

9. The apparatus according to claim 8, wherein said digital signal processor, said bus system, said program memory, said working memory, and said input/output unit are formed as an integrated circuit.