Voice recognition with parallel gender and age normalization

• US 8,010,358 B2
• Filed: 02/21/2006
• Issued: 08/30/2011
• Est. Priority Date: 02/21/2006
• Status: Active Grant

• Current Assignee: Sony Interactive Entertainment Inc. (Sony Group Corp.)
• Original Assignee: Sony Computer Entertainment Incorporated (Sony Group Corp.)
• Inventors: Chen, Ruxin
• Primary Examiner(s): Wozniak; James S
• Assistant Examiner(s): GUERRA-ERAZO, EDGAR X
• Application Number:
  US11/358,272
• Publication Number:
  US 20070198261A1
• Time in Patent Office: 2,016 Days
• Field of Search: 704/250, 704/255, 704/256, 704/246, 704/270.1, 704/E15.009, 704/245, 704/207, 704/208, 704/256.2
• US Class Current: 704/246
• CPC Class Codes: G10L 15/065 Adaptation
• CPC Class Codes: G10L 15/32 Multiple recognisers used in sequence or in parallel; Score combination systems therefor, e.g. voting systems

Abstract

Methods and apparatus for voice recognition are disclosed. A voice signal is obtained and two or more voice recognition analyses are performed on the voice signal. Each voice recognition analysis uses a filter bank defined by a different maximum frequency and a different minimum frequency and wherein each voice recognition analysis produces a recognition probability r_i of recognition of one or more speech units, whereby there are two or more recognition probabilities r_i. The maximum frequency and the minimum frequency may be adjusted every time speech is windowed and analyzed. A final recognition probability P_f is determined based on the two or more recognition probabilities r_i.

Claims

1. A method for voice recognition, the method comprising:

obtaining a voice signal;

performing two or more voice recognition analyses in parallel on the voice signal, wherein each voice recognition analysis uses a filter bank defined by a different maximum frequency and a different minimum frequency and wherein each voice recognition analysis produces a recognition probability r_i of recognition of one or more speech units, whereby there are two or more recognition probabilities r_i, wherein each recognition probability r_i represents a probability that the voice signal matches a recognized unit of speech, wherein performing the two or more voice recognition analyses in parallel includes performing a first voice recognition analysis on the voice signal based on a first maximum frequency f_max and a first minimum frequency f_min;

wherein f_min and f_max are adjusted dynamically during the first voice recognition; and

performing one or more additional voice recognition analyses on the voice signal based on a different maximum frequency given by f_max±

Δ

f_max and a different minimum frequency given by f_min±

Δ

f_min,where Δ

f_max<

f_maxand Δ

f_min<

f_min; and

determining a final recognition probability P_f based on the two or more recognition probabilities r_i;

determining a confidence measure based on the final recognition probability and the two or more recognition probabilities r_i; and

using the confidence measure as an acceptance criterion for the final recognition probability P_f after determining the final recognition probability P_f.

2. The method of claim 1 wherein determining the final recognition probability includes selecting a maximum probability of the two or more recognition probabilities as the final recognition probability P_f.

3. The method of claim 1 wherein obtaining a voice signal includes over-sampling the voice signal at a sampling frequency that is greater than a working feature analysis frequency.

4. The method of claim 3 wherein the sampling frequency is greater than a training time speech sampling rate.

5. The method of claim 1 wherein f_min and f_max are adjusted for each time window of the first voice recognition.

6. The method of claim 1 wherein different voice recognition analyses use the same values of f_min and f_max.

7. The method of claim 1 wherein Δ

f_min is between about 1% and about 5% of f_min.

8. The method of claim 1 wherein Δ

f_max is between about 1% and about 5% of f_max.

9. The method of claim 1 wherein performing one or more additional voice recognition analyses includes:

performing a second voice recognition analysis based on a maximum frequency given by f_max−

Δ

f_max and a minimum frequency given by f_min−

Δ

f_min;

performing a third voice recognition analyses is based on a maximum frequency given by f_max+Δ

f_max and a minimum frequency given by f_min−

Δ

f_max;

performing a fourth voice recognition analysis based on a maximum frequency given by f_max−

Δ

f_max and a minimum frequency given by f_min+Δ

f_max; and

performing a fifth voice recognition analysis based on a maximum frequency given by f_max+Δ

f_max and a minimum frequency given by f_min+Δ

f_min.

10. The method of claim 9 wherein the first, second, third, fourth and fifth voice recognition analyses are performed in parallel.

11. The method of claim 1 wherein performing two or more voice recognition analyses on the voice signal includes performing one or more gender-specific and/or age-specific voice recognition analyses.

12. The method of claim 1 wherein performing two or more voice recognition analyses on the voice signal includes performing two or more voice recognition analyses having different speaker category assumptions.

13. The method of claim 12 wherein determining the final recognition probability includes weighting the final recognition probability according to one or more speaker category probabilities associated with the different speaker category assumptions.

14. The method of claim 11 wherein the final recognition probability P_f is determined by:

15. The method of claim 12 wherein performing two or more voice recognition analyses having different speaker category assumptions includes:

performing a first voice recognition based on an assumption that the voice signal is obtained from a male speaker;

performing a second voice recognition based on an assumption that the voice signal is obtained from a female speaker;

performing a third voice recognition based on an assumption that the voice signal is obtained from a child speaker.

16. The method of claim 1 wherein the confidence measure is a maximum of the two or more recognition probabilities r_i, divided by a sum of the two or more recognition probabilities r_i.

17. The method of claim 1 wherein performing two or more voice recognition analyses on the voice signal includes adjusting the maximum frequency and the minimum frequency every time speech is windowed and analyzed.

23. The method of claim 1, wherein using the confidence measure as an acceptance criterion for the final recognition probability P_f includes comparing the confidence measure to a threshold value and rejecting the final recognition probability P_f if the confidence measure is below the threshold value.

18. A voice recognition system, comprising:

an interface adapted to obtain a voice signal;

one or more processors coupled to the interface; and

a memory coupled to the interface and the processor, the memory having embodied therein a set of processor readable instructions for configured to implement a method for voice recognition, the processor readable instructions including;

one or more instructions that direct the one or more processors to perform two or more voice recognition analyses in parallel on the voice signal, wherein each voice recognition analysis uses a filter bank defined by a different maximum frequency and a different minimum frequency and wherein each voice recognition analysis produces a recognition probability r_i , of recognition of one or more speech units, whereby there are two or more recognition probabilities r_i, wherein each recognition probability r_i represents a probability that the voice signal matches a recognized unit of speech, wherein each recognition probability r_i represents a probability that the voice signal matches a recognized unit of speech, wherein performing the two or more voice recognition analyses in parallel includes performing a first voice recognition analysis on the voice signal based on a first maximum frequency f _max and a first minimum frequency f_min;

wherein f_min and f_max are adjusted dynamically during the first voice recognition; and

performing one or more additional voice recognition analyses on the voice signal based on a different maximum frequency given by f_max±

Δ

f_max and a different minimum frequency given by f_min±

Δ

f_min,where Δ

f_max <

f_max and Δ

f_min<

f_min; and

an instruction to determine a final recognition probability P_f based on the two or more recognition probabilities; and

an instruction to determine a confidence measure based on the final recognition probability and the two or more recognition probabilities r_i; and

an instruction to use the confidence measure as an acceptance criterion for the final recognition probability P_f after determining the final recognition probability P_f.

19. The system of claim 18 wherein the one or more processors include two or more processor elements configured to perform parallel processing.

20. The system of claim 19 wherein the two or more processor units include a power processor element and one or more synergistic processor elements.

21. The system of claim 19 wherein the one or more instructions that direct the one or more processors to perform two or more voice recognition analyses on the voice signal are configured to perform the two or more voice recognition analyses in parallel.

22. The system of claim 18 wherein one or more instructions that direct the one or more processors to perform two or more voice recognition analyses adjust the maximum frequency and the minimum every time speech is windowed and analyzed.