Phoneme discrimination method

• US 5,202,926 A
• Filed: 09/12/1991
• Issued: 04/13/1993
• Est. Priority Date: 09/13/1990
• Status: Expired due to Term

• Current Assignee: OKI Electric Industry Company Limited
• Original Assignee: OKI Electric Industry Company Limited
• Inventors: Miki, Kei
• Primary Examiner(s): KEMENY, EMANUEL
• Application Number:
  US07/757,964
• Time in Patent Office: 579 Days
• Field of Search: 381/36-43, 381/31
• US Class Current: 704/222
• CPC Class Codes: G10L 15/02 Feature extraction for speech recognition; Selection of recognition unit
• CPC Class Codes: G10L 2015/025 Phonemes, fenemes or fenones being the recognition units

Abstract

To improve the accuracy of phoneme discrimination, a first phoneme discrimination method composes analyzing, not only static information on a voice but also dynamic information on the voice, power variations of the voice as a power-change pattern (PCP) and spectrum parameter variations of the voice as a time spectrum parameter pattern (TSP). After rough clustering based on the PCP, detailed classification is conducted using the TSP so that the voice is hierarchically discriminated to obtain phoneme symbols. This invention also provides a second phoneme discrimination method. To improve the accuracy of phoneme discrimination by a recognition system for unspecified, independent speakers, plural spectrum parameter codebooks classified in advance depending on voice qualities are provided. After rough clustering based on the PCP, the inputted voice is subjected to detailed discrimination with reference to plural codebooks corresponding to the PCP and also to voice quality determination, thereby obtaining phoneme symbols.

Claims

1. A phoneme discrimination method comprising the steps of:

analyzing the frequency of an inputted voice signal to compute, for predetermined time frame intervals, voice spectrum parameters and voice powers indicative of the intensity of the inputted voice;

combining the voice powers of adjacent frames to provide a power-change pattern (PCP) indicative of the time variance of the voice powers;

combining the spectrum parameters of neighboring frames to provide a time spectrum pattern (TSP) indicative of the time variance of the spectrum parameters;

vector-quantizing the PCP to determine power-change pattern codes (PCP codes);

vector-quantizing the TSP using a TSP codebook corresponding to the PCP-VQ codes in the respective frames to obtain TSP codes; and

determining the probable phoneme symbol of each time frame interval via a correlation of phoneme symbols to PCP code and TSP code.

3. The method of claim 1 wherein said determining step includes using a table of data correlating phonemes to PCP code data and TSP code data.

4. The method of claim 1 wherein said steps of vector-quantizing include referencing one or more previously prepared codebooks.

5. The method of claim 1 wherein said step of vector-quantizing the PCP includes operating a PCP-VQ unit which references a previously-prepared PCP codebook.

6. The method of claim 5 wherein said step of vector-quantizing the TSP includes operating a TSP-VQ unit which references a previously-prepared TSP codebook.

7. The method of claim 1 wherein said step of vector-quantizing the TSP includes operating a TSP-VQ unit which references a previously-prepared TSP codebook.

8. The method of claim 1 wherein said PCP vector-quantization step is in accordance with the formula:

##EQU5## where the index m is an integer ranging from 1 to M, where d(P_i, Y^m) indicates the distance between the PCP P_i and the PCP Y^m of the power code number m, where argmin means to determine the power code number which makes said distance shortest, and where M is the size of a PCP codebook.

9. The method of claim 1 wherein said step of vector-quantizing the TSP includes selecting the TSP codebook corresponding to the PCP code from a set of TSP codebooks and vector quantizing using the codebook so selected.

10. The method of claim 9 wherein said step of vector-quantizing the TSP is in accordance with the following formula:

##EQU6## where the PCP code is C_i, the time spectrum pattern code is Z_i, where U(C_i)^r corresponds to the code C_i and is a TSP comprising (2k+1)*J elements, where r is a code number allotted to each TSP, and where R(C_i) means the size of the TSP codebook corresponding to the code C_i.

2. A phoneme discrimination method comprising the steps of:

analyzing the frequency of an inputted voice signal to obtain, for predetermined time frame intervals voice spectrum parameters and voice powers indicative of the intensity of the inputted voice;

combining the voice powers of adjacent frames to provide a power-change pattern (PCP) indicative of the time variance of the voice powers;

vector-quantizing the PCP to obtain power-change pattern VQ codes (PCP-VQ codes);

vector-quantizing the spectrum parameters using plural spectrum parameter codebooks to obtain plural spectrum parameter VQ codes and plural quantization errors (VQ errors);

selecting the spectrum parameter codebook to minimize the sum of the plural VQ errors, and selecting optimal spectrum parameter codes on the basis of the spectrum parameter codebook so selected; and

determining the probable phoneme symbol of each frame via a correlation of phoneme symbols to a selected codebook, an optimal spectrum parameter code and a PCP code.

11. The method of claim 2 wherein said determining step includes using a table of data correlating phonemes to the codebook, spectrum parameter code data, and PCP code data.

12. A phoneme discrimination unit comprising:

an acoustic analyzer configured to provide voice spectrum parameters and voice powers indicative of the intensity of inputted voice signals;

a power-change patter (PCP) generator coupled to receive said voice powers from the acoustic analyzer;

a PCP vector quantizing unit coupled to the PCP generator;

a time spectrum parameter (TSP) pattern generator coupled to receive said voice spectrum parameters from the acoustic analyzer;

a TSP vector quantizing unit coupled to receive inputs from said TSP pattern generator and said PCP vector quantizing unit;

a phoneme determining unit coupled to receive inputs from said TSP vector quantizing unit; and

an output from said phoneme determining unit.

13. The phoneme discriminating unit of claim 12 further comprising a PCP codebook coupled to said PCP vector quantizing unit.

14. The phoneme discriminating unit of claim 13 further comprising a set of TSP codebooks coupled to said TSP vector quantizing unit.

15. The phoneme discriminating unit of claim 12 further comprising a set of TSP codebooks coupled to said TSP vector quantizing unit.

16. A method of discriminating phonemes from an inputted voice signal comprising the steps of:

extracting voice powers and spectrum parameters from said inputted voice signal by an acoustic analysis process;

obtaining a power change pattern (PCP) code from said voice powers;

processing the spectrum parameters to produce a time spectrum pattern (TSP);

vector-quantizing the TSP to obtain a TSP code after selecting from among a plurality of TSP codebooks one TSP codebook corresponding to the PCP code; and

thereafter

selecting a phoneme as a function of the PCP code and the TSP code.

17. The method of claim 16 wherein said selecting step includes referring to a previously stored table.

18. The method of claim 16 wherein said step of obtaining a PCP code includes combining a plurality of said voice powers to provide a PCP and vector-quantizing the PCP, including referring to a previously stored PCP codebook.

19. The method of claim 16 wherein said TSP codebooks have been previously stored.

20. The method of claim 16 wherein said acoustic analysis process provides said voice powers and said spectrum parameters for respective time intervals, and the step of selecting a phoneme corresponds to said time intervals.