Acoustic event detection in polyphonic acoustic data

1. A computer-implemented method comprising:

inputting, by one or more processing units, polyphonic acoustic data into a trained neural network trained by labeled monophonic acoustic data, wherein the trained neural network is trained by:

inputting, by one or more processing units, the labeled monophonic acoustic data into the trained neural network; and

obtaining, by one or more processing units, the trained neural network by using the label vectors of the labeled monophonic acoustic data as the ground truth of a second output of the trained neural network;

obtaining, by one or more processing units, a first output from a hidden layer of the trained neural network; and

determining, by one or more processing units, at least one acoustic classification of the polyphonic acoustic data based on the first output and a feature dictionary learnt from the trained neural network.

2. The method of claim 1 , wherein the determining comprises:

representing, by one or more processing units, the first output as a non-negative sparse combination of the dictionary;

computing, by one or more processing units, a membership function of the first output based on the non-negative sparse combination of the dictionary; and

determining, by one more processing units, an event of an acoustic classification occurred in the polyphonic acoustic data in response to a membership value of the membership function being larger than an assigned threshold.

3. The method of claim 1 , wherein the feature dictionary is learnt by:

inputting, by one or more processing units, the monophonic acoustic data into the trained neural network;

obtaining, by one or more processing units, a third output from a hidden layer of the trained neural network; and

learning, by one or more processing units, the feature dictionary based on the third output.

4. The method of claim 3 , wherein the learning is used to optimize an object function for the at least one acoustic classification.

5. The method of claim 1 , wherein the labeled monophonic acoustic data are denoted as feature vectors corresponding to a sound segment.

6. The method of claim 1 , wherein the hidden layer is the last hidden layer.

7. The method of claim 1 , wherein the trained neural network is a Recurrent Neural Network (RNN) with Long Short Term Memory (LSTM) neurons.

8. A computer system, comprising:

a processor;

a computer-readable memory coupled to the processor, the memory comprising instructions that when executed by the processor perform actions of:

inputting polyphonic acoustic data into a trained neural network trained by labeled monophonic acoustic data, wherein the trained neural network is trained by:

inputting the labeled monophonic acoustic data into the trained neural network; and

obtaining the trained neural network by using the label vectors of the labeled monophonic acoustic data as the ground truth of a second output of the trained neural network;

obtaining a first output from a hidden layer of the trained neural network; and

determining at least one acoustic classification of the polyphonic acoustic data based on the first output and a feature dictionary learnt from the trained neural network.

9. The system of claim 8 , wherein the determining comprises:

representing the first output as a non-negative sparse combination of the dictionary;

computing a membership function of the first output based on the non-negative sparse combination of the dictionary; and

determining an event of an acoustic classification occurred in the polyphonic acoustic data in response to a membership value of the membership function being larger than an assigned threshold.

10. The system of claim 8 , wherein the feature dictionary is learnt by:

inputting the monophonic acoustic data into the trained neural network;

obtaining a third output from a hidden layer of the trained neural network; and

learning the feature dictionary based on the third output.

11. The system of claim 10 , wherein the learning is to optimize an object function for the at least one acoustic classification.

12. The system of claim 8 , wherein the labeled monophonic acoustic data are denoted as feature vectors corresponding to a sound segment.

13. The system of claim 8 , wherein the hidden layer is the last hidden layer.

14. The system of claim 8 , wherein the trained neural network is a Recurrent Neural Network (RNN) with Long Short Term Memory (LSTM) neurons.

15. A computer program product, comprising a computer readable non-transitory storage medium having program instructions embodied therewith, the program instructions being executable by a processor to perform a method comprising:

inputting polyphonic acoustic data into a trained neural network trained by labeled monophonic acoustic data, wherein the trained neural network is trained by:

inputting the labeled monophonic acoustic data into the trained neural network; and

obtaining the trained neural network by using the label vectors of the labeled monophonic acoustic data as the ground truth of a second output of the trained neural network;

obtaining a first output from a hidden layer of the trained neural network; and

determining at least one acoustic classification of the polyphonic acoustic data based on the first output and a feature dictionary learnt from the trained neural network.

16. The computer program product of claim 15 , wherein the determining comprises:

representing the first output as a non-negative sparse combination of the dictionary;

computing a membership function of the first output based on the non-negative sparse combination of the dictionary; and

determining an event of an acoustic classification occurred in the polyphonic acoustic data in response to a membership value of the membership function being larger than an assigned threshold.

17. The computer program product of claim 15 , wherein the feature dictionary is learnt by:

inputting the monophonic acoustic data into the trained neural network;

obtaining a third output from a hidden layer of the trained neural network; and

learning the feature dictionary based on the third output.