Non-linear observation model for removing noise from corrupted signals
View Patent ↗A new statistical model describes the corruption of spectral features caused by additive noise. In particular, the model explicitly represents the effect of unknown phase together with the unobserved clean signal and noise. Development of the model has realized three techniques for reducing noise in a noisy signal as a function of the model.
1. A method for reducing noise in a noisy input signal, the method comprising:
converting a frame of the noisy input signal into an input feature vector; and
obtaining a noise-reduced feature vector by using an equation of the form
{circumflex over (x)}=y +ln|1 −e n−y |
where y is the input feature vector and n comprises a noise estimate.
2. A computer-readable medium having computer-executable instructions comprising:
converting a frame of the noisy input signal into an input feature vector; and
obtaining a noise-reduced feature vector by using an equation of the form
{circumflex over (x)}=y +ln|1 −e n−y |
where y is the input feature vector and n comprises a noise estimate.
3. A computer-readable medium having computer-executable instructions comprising:
a first module converting a frame of a noisy input signal into an input feature vector; and
a second module obtaining an estimate of a noise-reduced feature vector using an acoustic model of the acoustic environment, the acoustic model being based on a non-linear function that describes a relationship between the input feature vector, a clean feature vector, a noise feature vector and a phase relationship indicative of mixing of the clean feature vector and the noise feature vector, the phase relationship including a phase factor with a statistical distribution comprising a zero-mean Gaussian distribution, and wherein the acoustic model is based on a rotation of a coordinate space comprising the input feature vector, the clean feature vector and the noise feature vector to a change of variables having two dimensions.
4. The computer-readable medium of claim 3 wherein one dimension of the acoustic model is approximately Gaussian.
5. The computer-readable medium of claim 4 wherein the second module includes instructions to obtain the estimate of the noise-reduced feature vector as a function of performing a symbolic integration over one dimension of the acoustic model.
6. The computer-readable medium of wherein claim 4 the second module includes instructions to obtain the estimate of the noise-reduced feature vector as a function of performing a numeric integration over a second dimension of the acoustic model.
7. The computer-readable medium of claim 3 wherein the second module includes instructions for obtaining an estimate of the noise-reduced feature vector as a function of a Taylor series approximation.
8. The computer-readable medium of claim 7 wherein the second module includes instructions for obtaining an estimate of the noise-reduced feature vector as a function of a prior estimate of the noise-reduced feature vector.
9. The computer-readable medium of claim 8 wherein the second module includes instructions for obtaining an estimate of the noise-reduced feature vector as a function of repeatedly using an immediate prior estimate of the noise-reduced feature vector in a subsequent calculation for a selected number of iterations.
10. A method for reducing noise in a noisy input signal, the method comprising:
converting a frame of the noisy input signal into an input feature vector;
obtaining an acoustic model of the acoustic environment, the acoustic model being based on a non-linear function that describes a relationship between the input feature vector, a clean feature vector, a noise feature vector and a phase relationship indicative of mixing of the clean feature vector and the noise feature vector, the phase relationship being in the same domain as the clean feature vector and the noise feature vector, wherein obtaining the acoustic model includes modeling the phase relationship with a phase factor having a statistical distribution comprising a zero-mean Gaussian distribution, and wherein obtaining the acoustic model includes rotating a coordinate space of the acoustic model comprising the input feature vector, the clean feature vector and the noise feature vector to obtain a change of variables having two dimensions; and
using the input feature vector, the noise feature vector and the acoustic model to estimate the noise-reduced feature vector.
11. The method of claim 10 wherein one dimension of the acoustic model is approximately Gaussian.
12. The method of claim 11 wherein using the input feature vector, the noise feature vector and the acoustic model to estimate the noise-reduced feature vector includes integrating symbolically over one dimension of the acoustic model.
13. The method of claim 12 wherein using the input feature vector, the noise feature vector and the acoustic model to estimate the noise-reduced feature vector includes integrating numerically over a second dimension of the acoustic model.
14. The method of claim 12 wherein using the input feature vector, the noise feature vector and the acoustic model to estimate the noise-reduced feature vector includes using a Taylor series approximation.
15. The method of claim 14 wherein using the input feature vector, the noise feature vector and the acoustic model to estimate the noise-reduced feature vector includes using a prior estimate of the noise-reduced feature vector.
16. The method of claim 13 wherein using the input feature vector, the noise feature vector and the acoustic model to estimate the noise-reduced feature vector includes repeatedly using an immediate prior estimate of the noise-reduced feature vector in a subsequent calculation for a selected number of iterations.