Method for training a neural network and a data processing device
A computer-implemented method for training a neural network by using a data processing device is presented. The method comprises providing pairs of target audio data sets and distorted audio data sets to the neural network, providing values of user preference settings, wherein the values are associated to the pairs of target and distorted audio data sets, such that the neural network can be trained by a combination of the pairs of the target audio data and the distorted audio data in combination with the values of the user preference settings, wherein the distorted audio data and the values of the user preference settings are used as inputs to the neural network during training, and wherein a loss function used for training the neural network is a function of the values of the user preference settings, the target audio data, and the distorted audio data thereby providing for that a variety of user preference settings can be met with one and the same trained neural network.
1 . A computer-implemented method for training a neural network by using a data processing device, wherein the neural network is configured to provide, during inference, processed audio data based on audio input data, the method comprising:
inputting pairs of target audio data sets and distorted audio data sets to the neural network from the data processing device;
providing values of user preference settings from the data processing device;
associating the provided values of the user preference settings to the pairs of target and distorted audio data sets;
training the neural network by a combination of the pairs of the target audio data and the distorted audio data in combination with the provided values of the user preference settings,
wherein the training includes inputting the distorted audio data and the provided values of the user preference settings to the neural network, and
forming a training loop by inputting into a loss function used for training the neural network the values of the user preference settings, the target audio data, and processed audio data output from the neural network, the neural network corresponding to a communication device, to output loss data that is input to the trained neural network.
2 . The method according to claim 1 , wherein the values of the user preference settings comprise one or more parameter values for controlling denoising, dereverberation, and echo cancellation.
3 . The method according to claim 1 , further comprising:
inputting information about a type of audio output device intended to output, during inference, the processed audio data,
wherein the type of audio output device is associated to the pairs of target and distorted data sets.
4 . The method according to claim 3 , wherein the type of audio output device is a two-way communication device comprising a microphone and a speaker.
5 . The method according to claim 3 , wherein the type of audio output device is selected from a group comprising of a hearing aid, a headset, a sound bar, a loudspeaker, a speakerphone, a hearable, and a conference speaker.
6 . A method for configuring a communication device, said method comprising:
training a neural network according to the method of claim 1 ;
integrating the trained neural network into the communication device; and
adjusting values of the user preference settings such that processed audio data output from the communication device meets a quality of experience score.
7 . The method according to claim 6 , wherein the quality of experience score is indicative of one or more distortion types.
8 . The method according to claim 6 , wherein the communication device is a two-way communication device comprising:
a microphone; and
a speaker,
wherein a first digital signal processing device is associated with the microphone and a second DSP device is associated with the speaker, the method further comprising:
adjusting values of the user preference settings and the neural network, so that processed audio data from the first and second DSP device meets the quality of experience score.
9 . The method according to claim 6 , wherein the method is performed as part of a tuning process.
10 . A communication device provided with a neural network, wherein the neural network is configured to output, during inference, processed audio data based on audio input data, wherein the neural network is obtained by a process according to claim 1 .
11 . The communication device according to claim 10 , wherein the communication device is a two-way communication device comprising a microphone and a speaker.
12 . The communication device according to claim 10 , wherein the communication device is selected from a group comprising of a hearing aid, a headset, a sound bar, a loudspeaker, a speaker phone, and a conference speaker.
13 . A non-transitory computer-readable storage medium storing one or more programs configured to be executed by one or more processors of a processing system, the one or more programs comprising instructions for performing the method according to claim 1 .
14 . A data processing device for training a neural network that corresponds to a communication device, wherein the neural network is configured to provide, during inference, processed audio data based on audio input data, the data processing device comprising:
circuitry configured to:
execute an audio data obtaining function configured to obtain pairs of target audio data and distorted audio data,
execute a user preference obtaining function configured to obtain values of user preference settings, wherein the values are associated to the pairs of target and distorted data sets, and
train the neural network by inputting into a loss function the target audio data sets and the processed audio data sets output from the neural network, and the values of the user settings values, and outputting loss data from the loss function into the trained neural network.
15 . The data processing device according to claim 14 , wherein the values of the user preference settings comprise one or more parameters for controlling denoising, dereverberation, and echo cancellation.