AUTOMATIC SPEECH RECOGNITION METHOD AND APPARATUS, USING NON-LINEAR ENVELOPE DETECTION OF SIGNAL POWER SPECTRA

• US 20080288253A1
• Filed: 05/18/2007
• Published: 11/20/2008
• Est. Priority Date: 05/18/2007
• Status: Active Grant

• Current Assignee: STMicroelectronics SRL (STMicroelectronics NV)
• Original Assignee: STMicroelectronics SRL (STMicroelectronics NV)
• Inventors: Arnone, Luigi, Savi, Alberto Amilcare
• Granted Patent: US 7,877,252 B2
• Time in Patent Office: Days
• US Class Current: 704/251
• CPC Class Codes: G10L 15/02 Feature extraction for speech recognition; Selection of recognition unit
• CPC Class Codes: G10L 19/02 using spectral analysis, e.g. transform vocoders or subband vocoders

Abstract

An automatic speech recognition method includes converting an acoustic signal into a digital signal; determine a power spectrum of at least one portion of the digital signal; and non-linearly determining envelope values of the power spectrum at a plurality of respective frequencies, based on a combination of the power spectrum with a filter function. Non-linearly determining envelope values involves calculating each envelope value based on a respective number of values of the power spectrum and of the filter function and the respective number of values is correlated to the respective frequency of the envelope value.

Claims

1. An automatic speech recognition method, comprising:

converting an acoustic signal into a digital signal;

determining a power spectrum of at least one portion of the digital signal; and

non-linearly determining envelope values of the power spectrum at a plurality of respective frequencies, based on a combination of the power spectrum with a filter function;

wherein non-linearly determining envelope values comprises calculating each envelope value based on a respective number of values of the power spectrum and of the filter function and said respective number of values is correlated to the respective frequency of the envelope value.

2. A method according to claim 1 wherein the number of values is correlated to the frequency such that, if a first frequency associated with a first envelope value Y(k₁) is greater than a second frequency associated with a second envelope value Y(k₂), the number of values for calculating the first envelope value Y(k₁) is greater than number of values for calculating the second envelope value Y(k₂).

3. A method according to claim 1 wherein envelope values are calculated according to the following equations:

Y(k)=max_i[S_M(m_i/2)H(k−

m_i)]

m_i=π

A/2(i+A),i=0,1, . . . , k

where Y(k) are the envelope values, S_M(m_i) is the power spectrum, H(k−

m_i) is the filter function and A is a tuning parameter.

4. A method according to claim 1 wherein envelope values are calculated according to the following equation:

Y(k)=max_i[S_M(i)h(i)],i≦

k

where Y(k) are the envelope values, S_M(i) is the power spectrum and H(k−

m_i) is the filter function.

5. A method according to claim 1, wherein the filter function is a low-pass filter function.

6. A method according to claim 5 wherein the filter function is a cosine filter function.

7. A method according to claim 1, further comprising extracting a set of Mel Frequency Cepstral Coefficients associated with the least one portion of the digital signal, based on the envelope values.

8. A method according to claim 7 wherein extracting a set of Mel Frequency Cepstral Coefficients comprises filtering the envelope values with a MEL filterbank.

9. A method according to claim 1, further comprising dividing the digital signal into a plurality of signal frames.

10. A method according to claim 9, further comprising window-filtering the signal frames.

11. An automatic speech recognition apparatus comprising:

an input module for receiving an acoustic signal and converting the acoustic signal into a digital signal;

a spectral power estimation block for determining a power spectrum of at least one portion of the digital signal; and

a smoothing block for non-linearly determining envelope values of the power spectrum at a plurality of respective frequencies, based on a combination of the power spectrum with a filter function;

wherein the smoothing block is configured to calculate each envelope value based on a respective number of values of the power spectrum and of the filter function, said respective number of values being correlated to the respective frequency of the envelope value.

12. An apparatus according to claim 11 wherein the smoothing block is configured to correlate the number of values to the frequency such that, if a first frequency associated with a first envelope value Y(k₁) is greater than a second frequency associated with a second envelope value Y(k₂), the number of values for calculating the first envelope value Y(k₁) is greater than number of values for calculating the second envelope value Y(k₂).

13. An apparatus according to claim 11 wherein envelope values are calculated according to the following equations:

Y(k)=max_i[S_M(m_i/2)H(k−

m_i)]

vm_i=π

A/2(i+A),i=0,1, . . . , k

where Y(k) are the envelope values, S_M(m_i) is the power spectrum, H(k−

m_i) is the filter function and A is a tuning parameter.

14. An apparatus according to claim 11 wherein envelope values are calculated according to the following equation:

Y(k)=max_i[S_M(i)h(i)],i≦

k

where Y(k) are the envelope values, S_M(i) is the power spectrum and H(k−

m_i) is the filter function.

15. An apparatus according to claim 11, further comprising a MEL filterbank for filtering the envelope values.

16. A vocal control system comprising:

an automatic speech recognition apparatus for providing output messages in response to vocal commands imparted by a user, the automatic speech recognition apparatus comprising;

an input module for receiving an acoustic signal and converting the acoustic signal into a digital signal;

a spectral power estimation block for determining a power spectrum of at least one portion of the digital signal; and

a smoothing block for non-linearly determining envelope values of the power spectrum at a plurality of respective frequencies, based on a combination of the power spectrum with a filter function;

wherein the smoothing block is configured to calculate each envelope value based on a respective number of values of the power spectrum and of the filter function, said respective number of values being correlated to the respective frequency of the envelope value.

17. A system according to claim 16, further comprising:

a control unit; and

a peripheral unit controlled by the control unit, wherein the control unit is coupled to the automatic speech recognition apparatus for receiving output messages and for producing control signals for the peripheral unit, based on the output messages.

18. The vocal control system of claim 17 wherein the peripheral unit is a unit of a portable telephone.

19. The vocal control system of claim 17 wherein the peripheral unit is a unit of a vehicle.

20. The vocal control system of claim 17 wherein the peripheral unit is a unit of a household appliance.

21. A computer-readable medium containing instructions that, when executed, performs a method comprising:

determining a power spectrum of at least one portion of a digital signal; and

non-linearly determining envelope values of the power spectrum at a plurality of respective frequencies, based on a combination of the power spectrum with a filter function, the non-linearly determining envelope values of the power spectrum at a plurality of respective frequencies including calculating each envelope value based on a respective number of values of the power spectrum and of the filter function and said respective number of values is correlated to the respective frequency of the envelope value.

22. The computer-readable medium of claim 21 wherein the number of values is correlated to the frequency such that, if a first frequency associated with a first envelope value Y(k₁) is greater than a second frequency associated with a second envelope value Y(k₂), the number of values for calculating the first envelope value Y(k₁) is greater than number of values for calculating the second envelope value Y(k₂).

23. The computer-readable medium of claim 21 wherein envelope values are calculated according to the following equations:

Y(k)=max_i[S_M(m_i/2)H(k−

m_i)]

m_i=π

A/2(i+A),i=0,1, . . . , k

where Y(k) are the envelope values, S_M(m_i) is the power spectrum, H(k−

m_i) is the filter function and A is a tuning parameter.

24. The computer-readable medium of claim 21 wherein envelope values are calculated according to the following equation:

Y(k)=max_i[S_M(i)h(i)],i≦

k

where Y(k) are the envelope values, S_M(i) is the power spectrum and H(k−

m_i) is the filter function.

25-41. -41. (canceled)