Enhancing audio signals using sub-band deep neural networks
Systems and methods for enhancing reverberated audio signals are disclosed. In one embodiment, a method is disclosed comprising receiving an audio signal; partitioning a frequency domain representation of the audio signal into a plurality of sub-band vectors; inputting each sub-band vector into a corresponding deep neural network; calculating, using the corresponding deep neural networks, a plurality of output vectors for each sub-band; concatenating the plurality of output vectors to generate a clean audio feature matrix; and converting the clean audio feature matrix into a time-domain audio signal.
1. A method comprising:
receiving an audio signal;
partitioning a frequency domain representation of the audio signal into a plurality of sub-band vectors;
inputting each sub-band vector into a corresponding deep neural network;
calculating, using the corresponding deep neural networks, a plurality of output vectors for each sub-band vector;
generating at least one fully connected layer using a first subset of the output vectors;
concatenating a second subset of the output vectors with the fully connected layer to generate a clean audio feature matrix; and
converting the clean audio feature matrix into a time-domain audio signal.
2. The method of claim 1 , the generating at least one fully connected layer comprising generating at least two fully connected layers, the two fully connected layers comprising a first fully connected layer and a second fully connected layer.
3. The method of claim 2 , at least one output vector being used to generate both the first fully connected layer and second fully connected layer.
4. The method of claim 1 , the inputting each sub-band vector into the corresponding deep neural network comprising processing each sub-band vector using a first stage of the corresponding deep neural networks.
5. The method of claim 4 , the first stage of the corresponding deep neural networks comprising one or more hidden layers of the corresponding deep neural networks.
6. The method of claim 5 , the first stage of the corresponding deep neural networks additionally comprising one or more fully connected layers of the corresponding deep neural networks.
7. The method of claim 1 , the generating the least one fully connected layer comprising generating the fully connected layer using a first layer of a two-layer fully connected stage of the deep neural networks.
8. A non-transitory computer readable storage medium for tangibly storing computer program instructions capable of being executed by a computer processor, the computer program instructions defining the steps of:
receiving an audio signal;
partitioning a frequency domain representation of the audio signal into a plurality of sub-band vectors;
inputting each sub-band vector into a corresponding deep neural network;
calculating, using the corresponding deep neural networks, a plurality of output vectors for each sub-band vector;
generating at least one fully connected layer using a first subset of the output vectors;
concatenating a second subset of the output vectors with the fully connected layer to generate a clean audio feature matrix; and
converting the clean audio feature matrix into a time-domain audio signal.
9. The non-transitory computer readable storage medium of claim 8 , the generating at least one fully connected layer comprising generating at least two fully connected layers, the two fully connected layers comprising a first fully connected layer and a second fully connected layer.
10. The non-transitory computer readable storage medium of claim 9 , at least one output vector being used to generate both the first fully connected layer and second fully connected layer.
11. The non-transitory computer readable storage medium of claim 8 , the inputting each sub-band vector into the corresponding deep neural network comprising processing each sub-band vector using a first stage of the corresponding deep neural networks.
12. The non-transitory computer readable storage medium of claim 11 , the first stage of the corresponding deep neural networks comprising one or more hidden layers of the corresponding deep neural networks.
13. The non-transitory computer readable storage medium of claim 12 , the first stage of the corresponding deep neural networks additionally comprising one or more fully connected layers of the corresponding deep neural networks.
14. The non-transitory computer readable storage medium of claim 8 , the generating the least one fully connected layer comprising generating the fully connected layer using a first layer of a two-layer fully connected stage of the deep neural networks.
15. An apparatus comprising:
a processor; and
a storage medium for tangibly storing thereon program logic for execution by the processor, the stored program logic comprising:
logic, executed by the processor, for receiving an audio signal,
logic, executed by the processor, for partitioning a frequency domain representation of the audio signal into a plurality of sub-bands vectors,
logic, executed by the processor, for inputting each sub-band vector into a corresponding deep neural network,
logic, executed by the processor, for calculating, using the corresponding deep neural networks, a plurality of output vectors for each sub-band vector,
logic, executed by the processor, for generating at least one fully connected layer using a first subset of the output vectors,
logic, executed by the processor, for concatenating a second subset of the output vectors with the fully connected layer to generate a clean audio feature matrix, and
logic, executed by the processor, for converting the clean audio feature matrix into a time-domain audio signal.
16. The apparatus of claim 15 , the logic for generating at least one fully connected layer comprising logic, executed by the processor, for generating at least two fully connected layers, the two fully connected layers comprising a first fully connected layer and a second fully connected layer.
17. The apparatus of claim 16 , at least one output vector used to generate both the first fully connected layer and second fully connected layer.
18. The apparatus of claim 15 , the logic for inputting each sub-band vector into the corresponding deep neural network comprising logic, executed by the processor, for processing each sub-band vector using a first stage of the corresponding deep neural networks.
19. The apparatus of claim 18 , the first stage of the corresponding deep neural networks comprising one or more hidden layers of the corresponding deep neural networks.
20. The apparatus of claim 19 , the first stage of the corresponding deep neural networks additionally comprising one or more fully connected layers of the corresponding deep neural networks.
21. The apparatus of claim 15 , the logic for generating the least one fully connected layer comprising logic, executed by the processor, for generating the fully connected layer using a first layer of a two-layer fully connected stage of the deep neural networks.