Deep neural network based speech enhancement
A computer may segment a noisy audio signal into audio frames and execute a deep neural network (DNN) to estimate an instantaneous function of clean speech spectrum and noisy audio spectrum in the audio frame. This instantaneous function may correspond to a ratio of an a-priori signal to noise ratio (SNR) and an a-posteriori SNR of the audio frame. The computer may add estimated instantaneous function to the original noisy audio frame to output an enhanced speech audio frame.
1. A computer-implemented method comprising:
segmenting, by a computer, an audio signal into a plurality of audio frames;
generating, by the computer, a feature vector for an audio frame of the plurality of audio frames, the feature vector including values of a predetermined number of frequency channels of the audio frame and values of frequency channels of a predetermined number of audio frames on each side of the audio frame;
executing, by the computer, a deep neural network (DNN) on the feature vector to estimate an instantaneous function of a clean audio spectrum and a noisy audio spectrum of the audio frame; and
generating, by the computer, an enhanced speech audio frame corresponding to the audio frame based on noisy audio spectrum of the audio frame and the estimated instantaneous function of the clean audio spectrum and the noisy audio spectrum of the audio frame.
2. The computer-implemented method of claim 1 , further comprising:
outputting, by the computer, an enhanced speech audio signal corresponding to the audio signal and containing enhanced speech audio frame.
3. The computer-implemented method of claim 1 , further comprising:
estimating, by the computer, a frame-wise voice activity in association with estimating the instantaneous function of the clean audio spectrum and the noisy audio spectrum.
4. The computer-implemented method of claim 1 , wherein the instantaneous function of the clean audio spectrum and the noisy audio spectrum of the audio frame corresponds to a ratio of an a-priori signal-to-noise ratio (SNR) and an a-posteriori SNR of the audio frame.
5. The computer-implemented method of claim 1 , wherein the output layer of the DNN has a sigmoid activation function to constrain values in an output vector of the DNN between 0 and 1.
6. The computer-implemented method of claim 5 , further comprising:
estimating, by the computer, the instantaneous function by applying an inverse of the sigmoid activation function to the output vector of the DNN.
7. The computer-implemented method of claim 1 , wherein the step of generating the feature vector of the audio frame further comprises:
performing, by the computer, mean and variance normalization of corresponding values in the frequency channels of the audio frame.
8. The computer-implemented method of claim 1 , wherein the instantaneous function is a logarithmic ratio of the clean audio spectrum and the noisy audio spectrum of the audio frame.
9. The computer-implemented method of claim 1 , wherein the DNN is trained with a binary cross entropy loss function.
10. A system comprising:
a non-transitory storage medium storing a plurality of computer program instructions and a trained deep neural network (DNN); and
a processor electrically coupled to the non-transitory storage medium and configured to execute the plurality of computer program instructions to:
segment an audio signal into a plurality of audio frames;
generate a feature vector for an audio frame of the plurality of audio frames, the feature vector including a predetermined number of frequency channels of the audio frame and frequency channels of a predetermined number of audio frames on each side of the audio frame;
feed the feature vector to the DNN to estimate an instantaneous function of a clean audio spectrum and a noisy audio spectrum of the audio frame; and
generate an enhanced speech audio frame corresponding to the audio frame based on noisy audio spectrum of the audio frame and the estimated instantaneous function of the clean audio spectrum and the noisy audio spectrum of the audio frame.
11. The system of claim 10 , wherein the processor is configured to further execute the plurality of computer program instructions to:
output an enhanced speech audio signal corresponding to the audio signal and containing enhanced speech audio frame.
12. The system of claim 10 , wherein the processor is configured to further execute the plurality of computer program instructions to:
estimate a frame-wise voice activity in association with estimating the instantaneous function of the clean audio spectrum and the noisy audio spectrum.
13. The system of claim 10 , wherein the instantaneous function of the clean audio spectrum and the noisy audio spectrum of the audio frame correspond to a ratio of an a-priori signal-to-noise ratio (SNR) and an a-posteriori SNR of the audio frame.
14. The system of claim 10 , wherein the output layer of the DNN has a sigmoid activation function to constrain values in an output vector of the DNN between 0 and 1.
15. The system of claim 14 , wherein the processor is configured to further execute the computer program instructions to:
estimate the instantaneous function by applying an inverse of the sigmoid activation function to the output vector of the DNN.
16. The system of claim 10 , wherein the processor is configured to further execute the computer program instructions to:
perform mean and variance normalization of corresponding values in the frequency channels of the audio frame.
17. The system of claim 10 , wherein the instantaneous function is a logarithmic ratio of the clean audio spectrum and the noisy audio spectrum of the audio frame.
18. The system of claim 10 , wherein the DNN is trained with a binary cross entropy loss function.