Hearing device comprising a recurrent neural network and a method of processing an audio signal

A method of operating a hearing device, the method comprising

• providing at least one electric input signal in a time-frequency representation k, t, where k and t are frequency and time indices, respectively, and k represents a frequency channel, k=1, . . . , K, the at least one electric input signal being representative of sound and comprising target signal components and noise components;
• and
• providing an estimate of the target signal;
• providing an estimate of the noise;
• providing respective gain values in dependence of said target signal estimate and said noise estimate, wherein said gain estimator comprises a neural network, wherein the weights of the neural network have been trained with a plurality of training signals, and wherein the outputs of the neural network comprise real or complex valued gains, or separate real valued gains and real valued phases;
• providing that the neural network comprises at least one layer defined as a gated recurrent unit, the gated recurrent unit comprising memory in the form of a hidden state vector h, and wherein an output vector o(t) is provided by said gated recurrent unit in dependence of an input vector x(t) and said hidden state vector h(t−1), wherein said output o(t) at a given time step t is stored as said hidden state h(t) and used in the calculation of the output vector o(t+1) in the next time step t+1, and
• wherein the method further comprises
• determining changes Δx(i,t)=x(i,t)−{circumflex over (x)}(i,t−1), and Δh(j,t−1)=h(j,t−1)−ĥ(j,t−2) to the input vector x(t) and to the hidden state vector h(t−1), respectively, front one time instance, t−1, to the next, t, and where {circumflex over (x)}(i,t−1) and ĥ(j,t−2) are estimated values of x(i,t−1) and h(j,t−2), respectively, where indices i, j refers to the ith input neuron and the jth neuron of the hidden state, respectively, where 1≤i≤Nch,x and 1≤j≤Nch,oh, wherein Nch,x and Nch,oh is the number of processing channels of the input vector x and the hidden state vector h, respectively, and wherein the signal processor is further configured to provide that the number of updated channels among said Nch,x and said Nch,oh, processing channels of the modified gated recurrent unit for said input vector x(t) and said hidden state vector h(t−1), respectively, at said given time instance t is limited to a number of peak values Np,x, and Np,oh, respectively, where Np,x is smaller than Nch,x and Np,oh, is smaller than Nch,oh.

A method of operating an audio processing device comprising at least an input unit and a signal processor for processing outputs of the input unit and for providing a processed output, the signal processor comprising a neural network comprising at least one layer implemented as a modified gated recurrent unit (modified GRU) comprising memory in the form of a hidden state vector (h(t−1)), the method comprising

• providing by the input unit at least one electric input signal in a time-frequency representation k, t, where k and t are frequency and time indices, respectively, and k represents a frequency channel, k=1, . . . , K, the at least one electric input signal being representative of sound or image data;
• and
• providing an input vector x(t) to the at least one layer implemented as a gated recurrent unit (GRU) based on or comprising said at least one electric input signal, or a signal originating therefrom;
• calculating by the signal processor—at a given time instance t—changes Δx(i,t)=x(i,t)−{circumflex over (x)}(i,t−1), and Δh(j,t−1)=h(j,t−1)−ĥ(j,t−1)−ĥ(j,t−2) to the input vector x(t) and to the hidden state vector h(t−1), respectively, from one time instance, t−1, to the next, t, and where {circumflex over (x)}(i,t−1) and ĥ(j,t−2) are estimated values of x(i,t−1) and h(j,t−2), respectively, where indices i, j refers to the ith input neuron and the jth neuron of the hidden state, respectively, where 1≤i≤Nch,x and 1≤j≤Nch,oh, wherein Nch,x and Nch,oh is the number of processing channels of the input vector x and the hidden state vector h, respectively;
• and
• providing by the signal processor that the number of updated channels among said Nch,x and said Nch,oh processing channels of the modified gated recurrent unit for said input vector x(t) and said hidden state vector h(t−1), respectively, at said given time instance t is limited to a number of peak values Np,x, and Np,oh, respectively, where Np,x is smaller than and Nch,x, and Np,oh, is smaller than Nch, oh;
• calculating by the signal processor—at a given time instance—an output vector o(t) from the at least one layer implemented as a gated recurrent unit (GRU) in dependence of said input vector x(t) and said hidden state vector k(t−1), and wherein the output o(t) at a given time step t is stored as the hidden state h(t) and used in the calculation of the output vector o(t+1) in the next time step t+1;
• wherein the processed output is determined by the signal processor in dependence of said output vector o(t) or a signal originating therefrom;
• and
• wherein the processed output is used for controlling processing in the audio processing device and/or transmitted by the audio processing device to another device.