Noise suppression in a Mel-filtered spectral domain

1. A method comprising:

applying a window to a first representation of a speech signal in a time domain to provide a windowed representation of the speech signal;

converting the windowed representation of the speech signal in the time domain to a second representation of the speech signal in a frequency domain;

converting the second representation of the speech signal in the frequency domain to a third representation of the speech signal in a filtered spectral domain, wherein the third representation of the speech signal in the filtered spectral domain includes a plurality of Mel coefficients;

performing, by one or more processors, a noise suppression operation with respect to the third representation of the speech signal in the filtered spectral domain to provide a noise-suppressed representation of the speech signal that includes a plurality of noise-suppressed coefficients, wherein the noise suppression operation comprises:

determining a mean frame energy of the third representation of the speech signal, the mean frame energy being equal to a sum of the plurality of Mel coefficients divided by a number of the plurality of Mel coefficients; and

for each Mel coefficient of the plurality of Mel coefficients that is less than a noise floor threshold, setting that Mel coefficient to be equal to the noise floor threshold, the noise floor threshold being equal to the mean frame energy multiplied by a designated constant that is less than one.

2. The method of claim 1 , further comprising:

performing a logarithmic operation with respect to the plurality of noise-suppressed Mel coefficients to provide a plurality of respective revised coefficients;

truncating the plurality of revised coefficients to provide a truncated plurality of coefficients that includes fewer than all of the plurality of revised coefficients to represent the speech signal; and

performing a discrete transform with respect to at least one of the plurality of revised coefficients to de-correlate the plurality of revised coefficients or the truncated plurality of coefficients to de-correlate the truncated plurality of coefficients.

3. The method of claim 2 , further comprising:

applying a low-quefrency bandpass exponential cepstral lifter to each coefficient of the truncated plurality of coefficients to provide a liftered representation of the speech signal.

4. The method of claim 2 , further comprising:

performing a derivative operation with respect to the truncated plurality of coefficients to provide a plurality of respective first-derivative coefficients;

performing another derivative operation with respect to the plurality of first-derivative coefficients to provide a plurality of respective second-derivative coefficients; and

combining the truncated plurality coefficients, the plurality of first-derivative coefficients, and the plurality of second-derivative coefficients to provide a combined plurality of coefficients that represents the speech.

5. The method of claim 1 ,

wherein performing the noise suppression operation comprises:

determining a spectral noise estimate regarding the third representation of the speech signal; and

determining a plurality of signal-to-noise ratios that corresponds to the plurality of respective Mel coefficients, each signal-to-noise ratio representing a relationship between the corresponding Mel coefficient and the spectral noise estimate.

6. The method of claim 5 , wherein determining the spectral noise estimate comprises:

determining the spectral noise estimate based on a running average of an initial subset of the plurality of Mel coefficients.

7. The method of claim 5 , wherein performing the noise suppression operation further comprises:

determining a plurality of gains that corresponds to the plurality of respective Mel coefficients; and

multiplying the plurality of gains and the plurality of respective Mel coefficients to provide a plurality of respective speech estimates that represents the speech;

wherein each gain is substantially equal to a fixed maximum gain if the corresponding signal-to-noise ratio is greater than an upper signal-to-noise threshold;

wherein each gain is substantially equal to a fixed minimum gain if the corresponding signal-to-noise ratio is less than a lower signal-to-noise threshold; and

wherein each gain is based on a polynomial function of the corresponding signal-to-noise ratio if the corresponding signal-to-noise ratio is less than the upper signal-to-noise threshold and greater than the lower signal-to-noise threshold.

8. The method of claim 7 , further comprising:

determining a mean frame energy with respect to the plurality of speech estimates, the mean frame energy being equal to a sum of the plurality of speech estimates divided by a number of the plurality of speech estimates; and

for each speech estimate of the plurality of speech estimates that is less than a noise floor threshold, setting that speech estimate to be equal to the noise floor threshold, the noise floor threshold being equal to the mean frame energy multiplied by a designated constant that is less than one.

9. An automatic speech recognition system comprising:

one or more processors; and

a memory containing program code, which, when executed by at least one of the one or more processors, is configured to perform operations, the operations comprising: applying a window to a first representation of a speech signal in a time domain to provide a windowed representation of the speech signal;

converting the windowed representation of the speech signal in the time domain to a second representation of the speech signal in a frequency domain, the conversion module further configured to convert the second representation of the speech signal in the frequency domain to a third representation of the speech signal in a filtered spectral domain, wherein the third representation of the speech signal in the filtered spectral domain includes a plurality of Mel coefficients;

performing a noise suppression operation with respect to the third representation of the speech signal in the filtered spectral domain to provide a noise-suppressed representation of the speech signal that includes a plurality of noise-suppressed coefficients, wherein the noise suppression operation comprises:

determining a mean frame energy of the third representation of the speech signal, the mean frame energy being equal to a sum of the plurality of Mel coefficients divided by a number of the plurality of Mel coefficients; and

updating each Mel coefficient of the plurality of Mel coefficients that is less than a noise floor threshold to be equal to the noise floor threshold, the noise floor threshold being equal to the mean frame energy multiplied by a designated constant that is less than one.

10. The automatic speech recognition system of claim 9 ,

the operations further comprising:

a spectral noise estimator configured to determine determining a spectral noise estimate regarding the third representation of the speech signal; and

determining a plurality of signal-to-noise ratios that corresponds to the plurality of respective Mel coefficients, each signal-to-noise ratio representing a relationship between the corresponding Mel coefficient and the spectral noise estimate.

11. The automatic speech recognition system of claim 10 , wherein the spectral noise estimate is based on a running average of an initial subset of the plurality of Mel coefficients.

12. The automatic speech recognition system of claim 10 , the operations further comprising:

determining a plurality of gains that corresponds to the plurality of respective Mel coefficients; and

multiplying the plurality of gains and the plurality of respective Mel coefficients to provide a plurality of respective speech estimates that represents the speech;

wherein each gain is substantially equal to a fixed maximum gain if the corresponding signal-to-noise ratio is greater than an upper signal-to-noise threshold;

wherein each gain is substantially equal to a fixed minimum gain if the corresponding signal-to-noise ratio is less than a lower signal-to-noise threshold; and

wherein each gain is based on a polynomial function of the corresponding signal-to-noise ratio if the corresponding signal-to-noise ratio is less than the upper signal-to-noise threshold and greater than the lower signal-to-noise threshold.

13. The automatic speech recognition system of claim 12 , the operations further comprising:

determining a mean frame energy with respect to the plurality of speech estimates, the mean frame energy being equal to a sum of the plurality of speech estimates divided by a number of the plurality of speech estimates; and

updating each speech estimate of the plurality of speech estimates that is less than a noise floor threshold to be equal to the noise floor threshold, the noise floor threshold being equal to the mean frame energy multiplied by a designated constant that is less than one.

14. A computer-readable storage device having computer program logic recorded thereon for enabling a processor-based system to perform noise suppression in a filtered spectral domain, the computer-readable storage device comprising:

a first program logic that enables the processor-based system to apply a window to a first representation of a speech signal in a time domain to provide a windowed representation of the speech signal;

a second program logic that enables the processor-based system to convert the windowed representation of the speech signal in the time domain to a second representation of the speech signal in a frequency domain;

a third program logic that enables the processor-based system to convert the second representation of the speech signal in the frequency domain to a third representation of the speech signal in the filtered spectral domain, wherein the third representation of the speech signal in the filtered spectral domain includes a plurality of Mel coefficients;

a fourth program logic that enables the processor-based system to perform a noise suppression operation with respect to the third representation of the speech signal in the filtered spectral domain to provide a noise-suppressed representation of the speech signal that includes a plurality of noise-suppressed coefficients, wherein the noise suppression operation comprises:

a fifth program logic that enables the processor-based system to determine a mean frame energy of the third representation of the speech signal, the mean frame energy being equal to a sum of the plurality of Mel coefficients divided by a number of the plurality of Mel coefficients; and

a sixth program logic that enables the processor-based system to update each Mel coefficient of the plurality of Mel coefficients that is less than a noise floor threshold to be equal to the noise floor threshold, the noise floor threshold being equal to the mean frame energy multiplied by a designated constant that is less than one.

15. The computer-readable storage device of claim 14 ,

wherein the fourth program logic comprises:

first logic that enables the processor-based system to determine a spectral noise estimate regarding the third representation of the speech signal; and

second logic that enables the processor-based system to determine a plurality of signal-to-noise ratios that corresponds to the plurality of respective Mel coefficients, each signal-to-noise ratio representing a relationship between the corresponding Mel coefficient and the spectral noise estimate.

16. The computer-readable storage device of claim 15 , wherein the spectral noise estimate is based on a running average of an initial subset of the plurality of Mel coefficients.

17. The computer-readable storage device of claim 15 , wherein the fourth program logic further comprises:

third logic that enables the processor-based system to determine a plurality of gains that corresponds to the plurality of respective Mel coefficients; and

fourth logic that enables the processor-based system to multiply the plurality of gains and the plurality of respective Mel coefficients to provide a plurality of respective speech estimates that represents the speech;

wherein each gain is substantially equal to a fixed maximum gain if the corresponding signal-to-noise ratio is greater than an upper signal-to-noise threshold;

wherein each gain is substantially equal to a fixed minimum gain if the corresponding signal-to-noise ratio is less than a lower signal-to-noise threshold; and

wherein each gain is based on a polynomial function of the corresponding signal-to-noise ratio if the corresponding signal-to-noise ratio is less than the upper signal-to-noise threshold and greater than the lower signal-to-noise threshold.

18. The computer-readable storage device of claim 17 , further comprising:

a seventh program logic that enables the processor-based system to determine a mean frame energy with respect to the plurality of speech estimates, the mean frame energy being equal to a sum of the plurality of speech estimates divided by a number of the plurality of speech estimates; and

an eighth program logic that enables the processor-based system to update each speech estimate of the plurality of speech estimates that is less than a noise floor threshold to be equal to the noise floor threshold, the noise floor threshold being equal to the mean frame energy multiplied by a designated constant that is less than one.

19. The automatic speech recognition system of claim 9 , the operations further comprising:

performing a logarithmic operation with respect to the plurality of noise-suppressed coefficients to provide a plurality of respective revised coefficients;

truncating the plurality of revised coefficients to provide a truncated plurality of coefficients that includes fewer than all of the plurality of revised coefficients to represent the speech signal; and

performing a discrete transform with respect to at least one of the plurality of revised coefficients to de-correlate the plurality of revised coefficients or the truncated plurality of coefficients to de-correlate the truncated plurality of coefficients.

20. The computer program storage device of claim 14 , further comprising:

a seventh program logic that enables the processor-based system to perform a logarithmic operation with respect to the plurality of noise-suppressed coefficients to provide a plurality of respective revised coefficients;

a eighth program logic that enables the processor-based system to truncate the plurality of revised coefficients to provide a truncated plurality of coefficients that includes fewer than all of the plurality of revised coefficients to represent the speech signal; and

a ninth program logic that enables the processor-based system to a discrete transform with respect to at least one of the plurality of revised coefficients to de-correlate the plurality of revised coefficients or the truncated plurality of coefficients to de-correlate the truncated plurality of coefficients.