Acoustic signal processing with neural network using amplitude, phase, and frequency
According to one embodiment, a signal generation device includes one or more processors. The processors convert an acoustic signal and output amplitude and phase at a plurality of frequencies. The processors, for each of a plurality of nodes of a hidden layer included in a neural network that treats the amplitude and the phase as input, obtain frequency based on a plurality of weights used in arithmetic operation of the node. The processors generate an acoustic signal based on the plurality of obtained frequencies and based on amplitude and phase corresponding to each of the plurality of nodes.
1. A signal generation device comprising:
one or more processors configured to:
convert an acoustic signal and output amplitude and phase at a plurality of frequencies;
for each of a plurality of nodes of a hidden layer included in a neural network that treats the amplitude and the phase as input, obtain a frequency based on a plurality of weights, each of the plurality of weights being multiplied with an output of a corresponding node in a previous layer; and
generate an acoustic signal based on the plurality of obtained frequencies and based on amplitude and phase corresponding to each of the plurality of nodes,
wherein the one or more processors obtain the frequency that is set for a node in the previous layer corresponding to a weight having a highest absolute value.
2. A signal generation device according to claim 1 , wherein the neural network is a complex-valued neural network that includes a layer for inputting and outputting complex numbers.
3. The signal generation device according to claim 2 , wherein the one or more processors are further configured to obtain, for a plurality of nodes present in a hidden layer for inputting and outputting the complex numbers, a frequency based on absolute values of the plurality of weights.
4. The signal generation device according to claim 1 , wherein the neural network is an acoustic model learnt in such a way that feature of an acoustic signal is input and posterior probability of each processing unit representing at least either a phoneme, or a syllable, or a character, or a word piece, or a word is output.
5. The signal generation device according to claim 1 , wherein the neural network is a denoising autoencoder learnt in such a way that feature of an acoustic signal is input and feature having noise eliminated therefrom is output.
6. A signal generation device according to claim 1 , wherein the one or more processors are further configured to:
generate, for each of the plurality of nodes, a signal based on the obtained frequency, corresponding amplitude, and corresponding phase, and
generate the acoustic signal by synthesizing a plurality of signals generated for the plurality of nodes.
7. A signal generation device according to claim 1 , wherein
the neural network includes a plurality of hidden layers; and
the one or more processors obtain the frequency from the hidden layer that is a layer close to an output layer or a layer close to an input layer among the plurality of hidden layers.
8. A signal generation device according to claim 1 , wherein
the neural network includes a plurality of hidden layers; and
the one or more processors
obtain the frequency from one or more hidden layers among the plurality of hidden layers; and
generate the acoustic signal for each of the one or more hidden layers.
9. A signal generation system comprising:
one or more processors configured to:
convert an acoustic signal and output amplitude and phase at a plurality of frequencies;
for each of a plurality of nodes of a hidden layer included in a neural network that treats the amplitude and the phase as input, obtain a frequency based on a plurality of weights, each of the plurality of weights being multiplied with an output of a corresponding node in a previous layer; and
generate an acoustic signal based on the plurality of obtained frequencies and based on amplitude and phase corresponding to each of the plurality of nodes,
wherein the one or more processors obtain the frequency that is set for a node in the previous layer corresponding to a weight having a highest absolute value.
10. A signal generation method comprising:
converting that includes converting an acoustic signal and outputting amplitude and phase at a plurality of frequencies;
obtaining, for each of a plurality of nodes of a hidden layer included in a neural network that treats the amplitude and the phase as input, a frequency based on a plurality of weights, each of the plurality of weights being multiplied with an output of a corresponding node in a previous layer; and
generating an acoustic signal based on the plurality of obtained frequencies and based on amplitude and phase corresponding to each of the plurality of nodes,
wherein the obtaining include obtaining the frequency that is set for a node in the previous layer corresponding to a weight having a highest absolute value.
11. A computer program product having a non-transitory computer readable medium including programmed instructions, wherein the instructions, when executed by a computer, cause the computer to perform:
converting an acoustic signal and output amplitude and phase at a plurality of frequencies;
for each of a plurality of nodes of a hidden layer included in a neural network that treats the amplitude and the phase as input, obtaining a frequency based on a plurality of weights, each of the plurality of weights being multiplied with an output of a corresponding node in a previous layer; and
generating an acoustic signal based on the plurality of obtained frequencies and based on amplitude and phase corresponding to each of the plurality of nodes,
wherein the obtaining include obtaining the frequency that is set for a node in the previous layer corresponding to a weight having a highest absolute value.