Continuous speech recognition apparatus

• US 4,677,673 A
• Filed: 12/21/1983
• Issued: 06/30/1987
• Est. Priority Date: 12/28/1982
• Status: Expired due to Fees

• Current Assignee: Tokyo Shibaura Denki Kabushiki Kaisha
• Original Assignee: Tokyo Shibaura Denki Kabushiki Kaisha
• Inventors: Ukita, Teruhiko, Nitta, Tsuneo, Watanabe, Sadakazu
• Primary Examiner(s): Kemeny, E. S. Matt
• Application Number:
  US06/563,755
• Time in Patent Office: 1,287 Days
• Field of Search: 381/41-44, 364/513.5
• US Class Current: 704/253
• CPC Class Codes: G10L 15/00 Speech recognition G10L17/00 takes precedence

Abstract

Continuous speech signal is recognized using "rough" and "detail" parameters derived from prestored reference speech and current unknown speech. The detail parameters are 16 spectral coefficients, the rough parameters 2 or 4 spectral coefficients representing the signal. A word interval detector decides segmentation based on rough parameter similarity.

Claims

1. A continuous speech recognition apparatus comprising:

speech signal generating means for generating a speech signal in accordance with input speech;

acoustic analyzing means connected to said speech signal generating means for extracting acoustic parameter patterns, each having a plurality of acoustic parameter components according to the speech signal from said speech signal generating means for each frame, said acoustic analyzing means including a first output port from which an integer M acoustic parameter components corresponding to said plurality of extracted acoustic parameter components are generated, and a second output port from which an integer N which is smaller than M, acoustic parameter components corresponding to said plurality of extracted acoustic parameter components are generated;

speech-interval detection means connected to said acoustic analyzing means for detecting the speech interval of the input speech;

first and second memory means in which rough and detail reference pattern data representative of a plurality of speed units are stored;

a first data generation circuit connected to said acoustic analyzing means and first memory means to calculate the similarity between rough feature parameter data having a predetermined number of frames of said N acoustic parameter components from said acoustic analyzing means and the plurality of rough reference pattern data stored in said first memory means and generates the frame data indicative of the frame position of the rough feature parameter data which gives the similarity of not lower than a preset value together with the reference pattern data which gives said similarity of not lower than the preset value;

a second data generation circuit connected to said acoustic analyzing means and second memory means to detect a plurality of subperiod to be obtained by dividing the speech interval detected by said speech-interval detection means on the basis of the frame data from said first data generation circuit and calculate the similarity between detail feature pattern data having a preset number of frames of said M acoustic parameter components and representing the speech unit in each subperiod and the detail reference pattern data specified by the reference pattern indicative data corresponding to said subperiod, thereby generating the subperiod data representative of each subperiod together with the reference pattern indicative data and similarity data with respect to each subperiod data, and

a recognition circuit which detects the sum of each similarity data relating to the subperiods in each subperiod series which constitutes said speed interval, and detects maximum one of the sums of similarity data on the basis of the output data from said second data generation circuit, thereby recognizing the speech unit data with respect to said detected subperiod series as the input speech.

2. A continuous speech recognition apparatus according to claim 1, wherein said first data generation circuit includes:

a buffer memory which sequentially receives the rough feature parameter data from said acoustic analyzing means and stores the rough characteristic parameter data of a predetermined number of frames; and

a data generator which detects the similarity between the rough feature parameter data of said predetermined frames stored in said buffer memory and the plurality of reference pattern data in said first memory means and generates the frame data indicative of the frame position of one of said predetermined frames of rough feature pattern data whenever said similarity becomes a value of not lower than a preset value, and at the same time which generates the reference pattern designation data corresponding to the rough reference pattern data which gave this similarity of not lower than the preset value.

3. A continuous speech recognition apparatus according to claim 2, wherein said second data generation circuit comprises:

third memory means for sequentially storing the detail feature parameter data from said acoustic analyzing means;

fourth and fifth memory means for sequentially storing the frame data and reference pattern instruction data from said first data generation circuit;

control means for selectively reading out the frame data from said fourth memory means and reading out the detail feature parameter data from the frame positions of said third memory means corresponding to the subperiod to be determined by said frame data readout, and for reading out from said fifth memory means the reference pattern instruction data corresponding to the frame data read out from said fourth memory means and reading out the corresponding detail reference pattern data from said second memory means on the basis of said reference pattern instruction data read out;

a similarity calculator to calculate the similarity between the detail reference pattern data read out from said second memory means and the detail feature parameter data of the plurality of frames read out from said third memory means; and

sixth memory means for storing the frame data read out from said fourth memory means, corresponding reference pattern instruction data from said fifth memory means, and corresponding similarity data from said similarity calculator.

4. A continuous speech recognition apparatus according to claim 3, wherein said control means includes:

a first control circuit to selectively read out the frame data and reference pattern instruction data from said fourth and fifth memory means, respectively;

a second control circuit to read out the detail feature parameter data of a predetermined number of frames from the corresponding frame positions in said third memory means on the basis of the frame data read out from said fourth memory means; and

a pattern register to store the detail feature parameter data of said predetermined number of frames from said third memory means.

5. A continuous speech recognition apparatus according to claim 2, wherein said speech interval detection means detects the speech interval on the basis of the rough reference parameter data from said acoustic analyzing means.

6. A continuous speech recognition apparatus according to claim 1, wherein said second data generation circuit comprises:

third memory means for sequentially storing the detail feature parameter data from said acoustic analyzing means;

fourth and fifth memory means for sequentially storing the frame data and reference pattern instruction data from said first data generation circuit;

control means for selectively reading out the frame data from said fourth memory means and reading out the detail feature parameter data from the frame positions of said third memory means corresponding to the subperiod to be determined by said frame data read out, and for reading out the reference pattern instruction data corresponding to the frame data read out from said fourth memory means from said fifth memory means and reading out the corresponding detail reference pattern data from said second memory means on the basis of said reference pattern instruction data read out;

a similarity calculator to calculate the similarity between the detail reference pattern data read out from said second memory means and the detail feature parameter data of the plurality of frames read out from said third memory means; and

sixth memory means for storing the frame data read out from said fourth memory means, corresponding reference pattern instruction data from said fifth memory means, and corresponding similarity data from said similarity calculator.

7. A continuous speech recognition apparatus according to claim 6, wherein said control means includes:

a first control circuit to selectively read out the frame data and reference pattern instruction data from said fourth and fifth memory means, respectively;

a second control circuit to read out the detail feature parameter data of a predetermined number of frames from the corresponding frame positions in said third memory means on the basis of the frame data read out from said fourth memory means; and

a pattern register to store the detail feature parameter data of said predetermined number of frames from said third memory means.

8. A continuous speech recognition apparatus according to claim 1, wherein said speech interval detection means detects the speech interval on the basis of the rough feature parameter data from said acoustic analyzing means.

9. A continuous speech recognition apparatus according to claim 2, wherein said second data generation circuit comprises:

third memory means for sequentially storing the detail feature parameter data, each having said preset number of frames of M parameter components from said acoustic analyzing means;

fourth and fifth memory means for sequentially storing the frame data and reference pattern instruction data from said first data generation circuit;

control means for selectively reading out the frame data from said fourth memory means and reading out the detail feature parameter data from the frame positions of said third memory means corresponding to the subperiod determined by said frame data read out, and for reading out from said fifth memory means the reference pattern instruction data corresponding to the frame data read out from said fourth memory means and reading out the corresponding detail reference pattern data from said second memory means on the basis of said reference pattern instruction data read out;

a multiple similarity calculator to calculate the similarity between the detail reference pattern data read out from said second memory means and the detail feature parameter data of the plurality of frames read out from said third memory means;

and sixth memory means for storing the frame data read out from said fourth memory means, corresponding reference pattern instruction data from said fifth memory means, and corresponding similarity data from said multiple similarity calculator.