Methods for the design of hearing systems, accessory devices, and hearing algorithm scenarios

Abstract

A hearing system, an accessory device, an agent, and a method for contextual design of a hearing algorithm for a hearing device are disclosed, the method comprising: initializing a model including a parameterized objective function; providing one or more operating parameters indicative of the hearing algorithm to the hearing device; obtaining operating data including corresponding input data and output data of the hearing device; obtaining evaluation data indicative of a user evaluation of the output data; determining one or more updated operating parameters based on the model, the operating data, and the evaluation data; and providing the updated operating parameters to the hearing device.

Description

Technical Field

[0001] The present invention relates to a hearing system, an accessory device and related methods, and more particularly to a method for situational design of a hearing algorithm of a hearing device. Background Art

[0002] Today’s approaches to hearing algorithm design have inherent limitations. While hearing device users may be delighted when they leave the hearing aid provider’s office with a new hearing device, some may later be disappointed if their hearing device cannot be fitted (personalized) due to unforeseen issues that occur in the field. For example, in a restaurant, the amplified screeching sound of cutlery on a plate can ruin a nice dinner. Summary of the invention

[0003] Therefore, improved methods for hearing algorithm design are needed.

[0004] Therefore, the present invention provides a hearing system comprising an accessory device and a hearing device, an accessory device and a method for contextual design of a hearing algorithm of a hearing device.

[0005] A method for contextual design of a hearing algorithm for a hearing device is disclosed, the method comprising: initializing a model including a parameterized objective function; providing one or more operating parameters indicative of the hearing algorithm to the hearing device; obtaining operating data including corresponding input data and output data of the hearing device; obtaining evaluation data indicative of a user evaluation of the output data; determining one or more updated operating parameters based on the model, the operating data, and the evaluation data; and providing the updated operating parameters to the hearing device.

[0006] Furthermore, an accessory device for a hearing system is disclosed, the accessory device comprising a memory module, a processor module and a wireless interface, wherein the accessory device is configured to perform any of the methods disclosed herein.

[0007] Also disclosed is a hearing system comprising a hearing device as disclosed herein and an accessory device.

[0008] An important advantage of the present invention is that the hearing algorithm is continuously updated (in real time or near real time) during normal use of the hearing device.

[0009] The present invention can provide an improved listening experience for hearing device users. In addition, an efficient and convenient way to configure one or more hearing device parameters of a hearing device is provided. For example, the contextual hearing algorithm design can save one or more time-consuming revisits to a hearing device fitter. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] The above and other features and advantages of the present invention will become apparent to those skilled in the art through the following detailed description of exemplary embodiments with reference to the accompanying drawings, in which:

[0011] Figure 1 schematically illustrates an exemplary hearing system with a contextual design of a hearing algorithm,

[0012] Figure 2 is a flowchart of an exemplary method according to the present invention, and

[0013] Figure 3 An exemplary implementation of a model in an agent of an accessory device is shown.

[0014] Reference numerals list

[0015] 2 Hearing System

[0016] 4 Accessory devices, smartphones

[0017] 6 Hearing devices

[0018] 8 Agent

[0019] 10 External Devices

[0020] 12 users

[0021] 14 Receiving Model

[0022] 18 TRY

[0023] 20 EXECUTE

[0024] 22 EVALUATE

[0025] 24 Adaptation (ADAPT)

[0026] 26. ASSUME

[0027] 30 Models

[0028] 32 First Model

[0029] 34 Second Model

[0030] 36 A first line / link between the first model and the second model for providing a first output from the second model to the first model

[0031] 38 A second line / link between the first model and the second model for providing a second output from the second model to the first model

[0032] 40 nodes

[0033] 42 nodes

[0034] 100 Methods for contextual design of hearing algorithms for hearing devices

[0035] 102 Initializing a model with a parameterized objective function

[0036] 104 Providing one or more operating parameters indicative of a hearing algorithm to the hearing device

[0037] 104A Provide hearing algorithm f or hearing algorithm identifier indicating f to the hearing device

[0038] 104B provides one or more hearing device parameters θ to the hearing device

[0039] 106 Obtaining operation data including corresponding input data x(t) and output data y(t) of the hearing device

[0040] 108 Obtaining evaluation data r(t) indicating user evaluation of output data

[0041] 110 Determine one or more updated operating parameters based on the model, the operating data, and the evaluation data

[0042] 110A Solve probabilistic reasoning problems

[0043] 110B Implementing message passing on factor graph implementations

[0044] 110C Minimize Function

[0045] 112 Provide updated operating parameters to the hearing device

[0046] 114 Return to get the operation data. DETAILED DESCRIPTION

[0047] In the following, various exemplary embodiments and details are described with reference to the accompanying drawings when relevant. It should be noted that the drawings may or may not be drawn to scale, and elements of similar structure or function are represented by the same reference numerals in all drawings. It should also be noted that the drawings are intended only to facilitate the description of the embodiments. They are not an exhaustive description of the invention or a limitation on the scope of the invention. In addition, the illustrated embodiments do not need to have all the aspects or advantages shown. The aspects or advantages described in conjunction with a specific embodiment are not necessarily limited to that embodiment, and may also be implemented in any other embodiment even if not shown as such or if not so explicitly described.

[0048] In the contextual design process, problems and solutions are developed simultaneously through in situ experimentation with the environment. The present invention proposes to provide the current best solution suggestion until it fails, and further develop the problem statement through intelligent in situ experimentation with the environment via an agent / accessory device, which in turn enables the improvement of the solution suggestion. Contextual learning requires that the complete design cycle be performed in real time or near real time. In the present invention, the experimentation process can be performed by an agent, which is represented by a smart watch (or smartphone) resident software program or application capable of communicating with the hearing device through a communication link (such as a low-power Bluetooth link), see also Figure 1 .

[0049] The present invention relates to a method for contextual design of a hearing algorithm for a hearing device, wherein the method comprises initializing a model. The model may comprise a parameterized objective function. In one or more exemplary methods, the model is a generative probability model. The model may be a hierarchical Bayesian dynamics system.

[0050] The method includes providing one or more operating parameters indicative of a hearing algorithm to a hearing device. In one or more exemplary methods / systems, the one or more operating parameters may include a hearing algorithm identifier and / or a set of hearing device parameters. The set of hearing device parameters may include and / or indicate filter coefficients and / or values ​​thereof. The set of hearing device parameters may include and / or indicate noise cancellation parameters and / or values ​​thereof. The set of hearing device parameters may include and / or indicate a compressor gain and / or values ​​thereof. The hearing algorithm identifier may include and / or indicate a hearing algorithm selected from a plurality of hearing algorithms.

[0051] The method comprises, for example, in the accessory device, obtaining operational data comprising corresponding input data and output data of the hearing device, for example, by receiving the input data and / or the output data from the hearing device and / or an intermediate relay device. The method comprises, for example, in the accessory device, obtaining operational data comprising corresponding input data and output data of the hearing device by receiving the input data from the hearing device, and determining the output data by applying a current hearing algorithm in the accessory device. Thus, the accessory device may derive the output data based on the input data, thereby saving transmission power in the hearing device.

[0052] The method comprises: obtaining, for example in an accessory device, evaluation data r indicating a user evaluation of output data. The evaluation data is aligned with the input data and the output data, for example providing one or more data tuples comprising corresponding input data, output data and evaluation data.

[0053] The method comprises determining one or more updated operating parameters based on the model, the operating data and the evaluation data, and then providing the updated operating parameters to the hearing device. In other words, the operating parameters of the hearing device are adapted according to the input data, the output data and the evaluation data, thereby providing a contextual hearing algorithm design, i.e. the operating parameters of the hearing device are advantageously adapted based on user evaluations during normal use of the hearing device.

[0054] The method comprises providing updated operating parameters to the hearing device. In one or more exemplary methods, providing the one or more operating parameters to the hearing device comprises providing a hearing algorithm identifier to the hearing device, for example by wirelessly transmitting the hearing algorithm identifier to the hearing device. Thereby, the agent is able to select a suitable or optimal hearing algorithm, for example from a plurality of hearing algorithms comprising a first hearing algorithm and a second hearing algorithm. In one or more exemplary methods, providing the one or more operating parameters to the hearing device comprises providing the one or more hearing device parameters to the hearing device, for example by wirelessly transmitting a set of hearing device parameters to the hearing device.

[0055] In one or more exemplary methods, the generative probability model is a generative probability model m given by:

[0056] p(x,y,r,f,θ,z|m),

[0057] Where x is the input data of the hearing device, y is the output data of the hearing device, r is the evaluation data, f is the hearing algorithm, θ is the hearing device parameter, and z is the latent variable.

[0058] In one or more exemplary methods, the hearing algorithm is a dynamical system given by:

[0059] s t+1 =f(s t ,x t ,θ)

[0060] y t =g(s t )

[0061] Where t is the time index; x t is the input data, s t is the state vector, y t As state s t , f(·) represents the hearing algorithm, and θ comprises a set of hearing device parameters.

[0062] Determining one or more updated operating parameters may include solving a probabilistic inference problem given one or more constraints (eg, the first constraint and / or the second constraint).

[0063] In one or more exemplary methods, determining one or more updated operating parameters includes solving a probabilistic inference problem given by

[0064] p(f,θ|d ＜t ={(x,y,r) ＜t},r ＞t >0,m),

[0065] Where f is the hearing algorithm, θ is the hearing device parameter, and d <t ={(x,y,r) <t} is the first constraint on past observations, and r >t >0 is a second constraint that future evaluation data should indicate positive user evaluations of the output data.

[0066] In one or more exemplary methods, solving the probabilistic reasoning problem includes performing message passing, for example, performing message passing on a factor graph realization of the model. The factor graph realization may be a Forney-type factor graph realization of the model.

[0067] In one or more exemplary methods, determining one or more updated operating parameters based on the model, the operating data, and the evaluation data includes minimizing a function, wherein the function is a divergence function, such as an alpha divergence function.

[0068] In one or more exemplary methods, the alpha divergence function is a function F given by:

[0069]

[0070] Where the function F is a function of the variational distribution q(z) and d is a data tuple (x,y,r). The function F[q] is called the variational free energy (VFE) functional. VFE is a distance measure between probability distributions p and q (see formula). VFE is an example of a larger class of distance measures (between probability distributions) called alpha divergence. When VFE is used as a distance measure, probabilistic reasoning in factor graphs leads to the "Variational Message Passing" (VMP) algorithm. The VFE and VMP algorithms are examples of how to perform reasoning. Other metrics / methods for reasoning based on minimizing alpha divergence can be used, such as expectation propagation and belief propagation.

[0071] In one or more exemplary methods, the model includes a first model and a second model. In one or more exemplary methods, the first model is a generative probabilistic model for input data and output data, and the second model is a generative probabilistic model for evaluation data. In one or more exemplary methods, the first model is based on the output of the second model, for example, Figure 3 u t (1) and u t (2) For example, the output of the second model may be for the hearing device to modulate or adapt one or more accuracy parameters in the first model. In other words, the first model and the second model are coupled models.

[0072] In addition, an accessory device for a hearing system is provided. The accessory device includes a memory module, a processor module, and a wireless interface. In one or more exemplary methods, the accessory device is configured to perform any of the methods disclosed herein. In addition, an agent is disclosed, wherein the agent is configured to perform any of the methods disclosed herein. The agent can be a software module, such as an application, that can be installed in the memory module of the accessory device.

[0073] Furthermore, a hearing system is disclosed, which comprises a hearing device and an accessory device.

[0074] In addition, a hearing device is disclosed. The hearing device may be a wearable device or a hearing aid, wherein the processor is configured to compensate for hearing loss of a user.

[0075] The hearing device may be of behind-the-ear (BTE) type, in-the-ear (ITE) type, in-the-canal (ITC) type, receiver-in-the-canal (RIC) type or receiver-in-the-ear (RITE) type. The hearing aid may be a binaural hearing aid.

[0076] The hearing device is configured for wireless communication with one or more devices, such as, for example, another hearing device that is part of a binaural hearing system, and / or with one or more accessory devices (such as a smartphone and / or a smart watch). The hearing device optionally includes an antenna for converting one or more wireless input signals (e.g., a first wireless input signal and / or a second wireless input signal) into an antenna output signal. The wireless input signal may originate from an external source, such as a spouse microphone device, a wireless TV audio transmitter, and / or a distributed microphone array associated with a wireless transmitter. The one or more wireless input signals may originate from, for example, another hearing device that is part of a binaural hearing system, and / or from one or more accessory devices.

[0077] The hearing device comprises a processor for processing an input signal, eg according to a hearing algorithm, eg to compensate for a hearing loss of a user of the hearing device.The processor provides an electrical output signal to the processor which is based on the input signal.

[0078] The hearing device is configured to receive one or more operating parameters of a hearing algorithm indicated to the hearing device, and to process an input signal according to the one or more operating parameters for providing an output signal. The output signal is converted into an audio output signal via a receiver. The audio output signal is output to a user. The hearing device is configured to wirelessly transmit the input signal (or one or more parameters or data representing it) to an accessory device / agent. In one or more exemplary hearing devices, the hearing device is configured to wirelessly transmit the output signal (or one or more parameters or data representing it) to the accessory device / agent. In one or more exemplary hearing devices, the user evaluates the output signal / hearing algorithm by providing user feedback via a button or other input interface (e.g., a motion sensor) on the hearing device. Therefore, the hearing device can be configured to determine and / or send evaluation data indicating a user evaluation of the output data / hearing algorithm to the accessory device / agent. The hearing device is configured to receive updated operating parameters of the hearing algorithm indicated to the hearing device, and to process the input signal according to the updated operating parameters for providing an output signal.

[0079] Figure 1 An exemplary hearing system is shown, which illustrates the contextual design of the hearing algorithm according to the present invention. The hearing system 2 comprises an accessory device 4, illustrated here as a smartphone, and a hearing device 6. The accessory device 4 is configured for wireless communication with the hearing device 6, e.g. via protocol or other wireless protocol and comprises or has installed an agent / application 8. Furthermore, the accessory device 4 is configured for wireless communication with a server device or other external device 10, for example for downloading the agent / application 8 and / or specific settings of the agent / application. A user 12 of the hearing device 6 provides user feedback on the operation / performance of the hearing device 6 to the agent 8 via an interface of the accessory device 4. In the hearing system 2, a model such as a generative probabilistic model including a parameterized objective function is initialized in the agent 8, for example by retrieving the model from a memory and / or receiving the model 14 from the external device 10. The model may be a model m given by p(x, y, r, f, θ, z|m), where x is input data of the hearing device, y is output data of the hearing device, r is evaluation data, f is a hearing device identifier, θ is a hearing device parameter, and z is a latent variable.

[0080] consider Figure 1 The agent-based hearing device design scheme is based on the assumption that the hearing aid algorithm is a dynamic system given by:

[0081] s t+1 =f(s t ,x t ,θ)

[0082] yt =g(s t )

[0083] Among them, t is the time mark; x t is the input data, s t is the state vector, y t As state s t , f(·) represents the hearing algorithm (equation), and θ comprises a set of hearing device parameters or adjustment parameters.

[0084] Listening algorithm design includes three tasks:

[0085] 1. (Algorithm design task). Specify the signal processing algorithm f(·).

[0086] 2. (Fitting task). Assign values ​​to the hearing device parameters θ.

[0087] 3. (Performance evaluation task). Evaluate the performance of selected values ​​of f(·) and θ.

[0088] Agent 8 passes data tuple D = {d t =(x,y,r) t :t=1,…,∞} observes its environment. Thus, the agent obtains operational data comprising corresponding input data x and output data y of the hearing device and evaluation data r indicating a user evaluation of the output data. Basically, this means that the agent 8 observes the current hearing device input and output data (x t ,y t ) and user evaluation (response) of the input and output data t .

[0089] In one or more exemplary systems / methods, the agent 8 obtains operational data comprising corresponding input data x and output data y of the hearing device by receiving only the input data x from the hearing device and then determining or calculating the corresponding output data y based on the agent's knowledge of the hearing algorithm executed by the hearing device. In one or more exemplary systems / methods, the agent 8 obtains operational data comprising corresponding input data x and output data y of the hearing device by receiving both the input data x and the output data y from the hearing device.

[0090] The user evaluation or response may include a button press (e.g., like, dislike), a gesture, or a physiological measurement such as heart rate or EEG signal. In one or more exemplary systems / methods, the evaluation data may be obtained via an interface of the hearing device (e.g., one or more buttons) and / or via a graphical user interface of an accessory device. In one or more exemplary systems / methods, the evaluation data may be obtained via a speech recognition module of an accessory device interface.

[0091] In one or more exemplary systems / methods described herein, t = 0 indicates a neutral or missing (not observed) performance evaluation, r t >0 involves positive evaluation and r t <0 corresponds to an unsatisfactory hearing device behavior. In other words, a good evaluation from the user corresponds to a positive value of r, such as r=1, a neutral evaluation from the user corresponds to r=0, and a negative evaluation from the user corresponds to a negative value of r, such as r=-1.

[0092] Furthermore, the agent 8 provides the updated operating parameters f, θ to the hearing device 6. In other words, the agent 8 is able to update the action A = {a t =(f,θ) t :t=1,…,∞} to the hearing device 6. In this case, the action A includes a suggestion or instruction for the signal processing algorithm f(.) and the hearing device parameter setting θ. The agent 8 (from the end user's perspective) should make the most interesting suggestion a with as little interaction (evaluation) as possible. t .

[0093] The method for contextual design of listening algorithms consists of four steps or cycles of tasks that are performed in situ and optionally repeated. The cycle includes the following tasks:

[0094] a try (TRY) 18, which comprises an act of providing an operating parameter to the hearing device;

[0095] EXECUTE 20, which includes the action of obtaining operation data (input data and / or output data) from the hearing device;

[0096] EVALUATE 22, which includes the act of obtaining evaluation data indicative of a user evaluation of the output data; and

[0097] Adapt (ADAPT) 24, which includes the act of determining one or more updated operating parameters based on the model, the operating data, and the evaluation data.

[0098] The method is seeded or initialized by pushing 26 the model to the agents (also referred to as hypothetical tasks).Exemplary design phases are described in further detail below.

[0099] Hypothetical task (reference 26)

[0100] A fully probabilistic modeling approach is adopted. This approach starts with the specification of a generative model:

[0101] p(x,y,r,f,θ,z|m) (1)

[0102] Here, m identifies the agent. We use unscripted variables such as x≡(x1,x2,…) to indicate sequences (and similarly for other variables). The generative model is a probabilistic joint model of all variables in the system under study. The model of the agent includes sensory / observed input signals (data tuples) d=(x,y,r), actions a=(f,θ), and internal (hidden, latent, unobserved) variables z. Here, we call this the hypothesis step, which reflects the assumption that the model m is true. In the context of hearing device signal processing, it is recommended that the model is a hierarchical Bayesian dynamics system, see Figure 1 Further details of the structure of the model will be described in more detail below.

[0103] Try the task (ref. 18)

[0104] If generate_model is true, the agent is configured so that actions are subject to the following constraints:

[0105] 1. We want to respect all past user observations of user preferences, i.e., we want to respect observation d ＜t ={(x,y,r) ＜t}.

[0106] 2. We hope that future user feedback will be positive (because we hope that users will be happy in the future), i.e., we hope that r ＞t >0.

[0107] This is the task to be attempted, which can be described as a probabilistic reasoning problem:

[0108] p(f,θ|d ＜t ={(x,y,r) ＜t},r ＞t >0,m) (2)

[0109] Equation 2 indicates that the computation that suggests interesting algorithms and parameter settings is posed as an inference problem. In particular, the posterior distribution with respect to (f,θ) under constraints (1) and (2) is computed. This Bayesian inference task (2) is typically intractable, but a method for computing an approximate solution to equation (2) is described below.

[0110] Execute the task (refer to 20)

[0111] The task involves executing the hearing algorithm specified by equation (3), which is implemented at the hearing device. The hearing algorithm is a dynamical system given by:

[0112] s t+1 =f(s t ,x t ,θ) (3)

[0113] y t =g(st )

[0114] Among them, t is the time mark; x t is the input data, s t is the state vector, y t As state s t , f(·) represents the hearing algorithm, and θ comprises a set of hearing device parameters.

[0115] Assessment Tasks (Ref. 22)

[0116] At any time, a hearing device user may respond by submitting a “response” signal r t The response signal may be a button press on the accessory device 4 or on an auxiliary accessory device, which forwards the evaluation data to the agent 8 to indicate that the user is dissatisfied with the current performance of the hearing aid. In one or more exemplary methods / systems, the evaluation data may include physiological data of the user, such as heart rate and / or EEG signals.

[0117] Thus, the agent 8 acquires signals (evaluation data) related to the performance of the currently active algorithm. As discussed, it is assumed that the agent is also provided with current input data and optionally output data indicating the behavior of the hearing device. The agent can determine the output data based on the input data, thereby reducing power consumption in the hearing device. In other words, the agent observes the data tuple d t =(x t ,y t ,r t ). Audio signals are typically sampled at a data rate of about 20K samples / second. For most user interfaces, the response signal is sampled at a much lower rate. In that case, assuming r t The sampling rate of the audio signal is upsampled in an appropriate manner. For example, in the case of a like / dislike button, the following value can be assigned to r:

[0118]

[0119] Adaptation tasks (ref. 24)

[0120] In this step, Agent 8’s model absorbs the following observations:

[0121]

[0122] In equation (5), the observed value Assigned to the model variable d t =(x t ,y t ,r t), and constrain any (counterfactual) future observations of the user response to be positive. Optionally, this is followed by a renormalization of the probability distribution. The effect of this replacement is that any subsequent (Bayesian) inference of the model will take into account all (factual and counterfactual) observations (which are constraints 1 and 2 described in the attempt step).

[0123] The try-execute-evaluate-adapt action can be repeated at all time steps.

[0124] The try-execute-evaluate-adapt action may be repeated at a certain frequency, for example, once every 1 minute, once every 5 minutes, once every 10 minutes, or once every 15 minutes. In one or more exemplary methods / systems, the try-execute-evaluate-adapt action may be triggered by a user evaluation and / or stopped in response to a user input. In one or more exemplary methods / systems, a loop of the try-execute-evaluate-adapt action may be executed when r≠0, i.e., whenever a user makes a relevant evaluation / feedback.

[0125] Figure 2 Flowchart of an exemplary method for contextual design of a hearing algorithm for a hearing device. The method 100, optionally performed in an accessory device, comprises: initializing 102 a model including a parameterized objective function; providing 104 to a hearing device one or more operating parameters indicating a hearing algorithm, the one or more operating parameters optionally comprising a hearing algorithm identifier and / or a set of hearing device parameters; acquiring 106 operating data including corresponding input data x(t) and output data y(t) of the hearing device; acquiring 108 evaluation data r(t) indicating a user evaluation of the output data; determining 110 one or more updated operating parameters based on the model, the operating data and the evaluation data; and providing 112 the updated operating parameters to the hearing device. The method returns 114 to acquiring 106 the operating data.

[0126] In one or more exemplary methods, providing 104 one or more operating parameters to the hearing device comprises providing 104A a hearing algorithm f or a hearing algorithm identifier to the hearing device. In other words, the agent 8 may be able to select which hearing algorithm the hearing device 6 should apply, for example from a set of multiple hearing algorithms.

[0127] In one or more exemplary methods, providing 104 one or more operating parameters to the hearing device includes providing one or more hearing device parameters θ to the hearing device 104B.

[0128] In method 100, the model is a generative probability model m given by:

[0129] p(x,y,r,f,θ,z|m),

[0130] Where x is the hearing device input data, y is the hearing device output data, r is the evaluation data, f is the hearing algorithm identifier, θ is the hearing device parameter, and z is the latent variable, and the hearing algorithm is a dynamical system given by:

[0131] s t+1 =f(s t ,x t ,θ)

[0132] y t =g(s t )

[0133] Among them, t is the time mark; x t is the input data, s t is the state vector, y t As state s t , f(·) represents the hearing algorithm, and θ comprises a set of hearing device parameters.

[0134] In one or more exemplary methods, determining 110 one or more updated operating parameters includes solving 110A a probabilistic inference problem, such as given by:

[0135] p(f,θ|d ＜t ={(x,y,r) ＜t},r ＞t >0,m),

[0136] Where f is the hearing algorithm, θ is the hearing device parameter, and d <t ={(x,y,r) <t} is the first constraint on past observations, and r >t >0 is a second constraint that future evaluation data should indicate positive user evaluations of output data.Solving 110A the probabilistic inference problem includes implementing 110B message passing on a factor graph implementation (eg, a Forney-type factor graph of the model).

[0137] In one or more exemplary methods, wherein determining 110 one or more updated operating parameters based on the model, operating data, and evaluation data includes minimizing 110 a C function, wherein the function is an alpha divergence function, such as a function F given by:

[0138]

[0139] Wherein, function F is a function of variational distribution q(z), and d is a data tuple (x, y, r). It should be understood that other divergence functions may be used.

[0140] Figure 3An exemplary implementation of the model is shown in an agent 8. Typically, an agent may be implemented as an application or software module installed in the memory of an accessory device such as a smartphone, smartwatch or tablet computer.

[0141] The structure of the model 30 of the agent 8 is shown inside the accessory device 4. The model 30 comprises a first model 32 and a second model 34, wherein the first model 32 is a generative probability model for the input data x(t) and the output data y(t), and the second model 34 is a generative probability model for the evaluation data r(t). Thus, the first model 32 is a generative probability model for the hearing device input (input data) and output (output data) signals. The second model 34 is a generative probability model for user responses Lines 36, 38 connect the states in the second model 34 (user model) to the factors in the first model 32. In summary, the agent's model 30 is a "sensory" input for the agent (i.e., for {d t =(x t ,y t ,r t ):t=1,2,…}).

[0142] The agent iteratively determines the best tuning parameter settings or operating parameters (θ) for a given hearing algorithm f. From a set of candidate algorithms {f k The task of finding the best hearing algorithm f among {:k=1,...K} is similar to the parameter estimation problem. The first model 32 is a (probabilistic) model of the hearing device signal processing algorithm. Typically, the model of the signal processing algorithm will be a dynamic hierarchical structure. A typical example of the bottom layer is given by:

[0143]

[0144]

[0145] in, is the internal state vector of the nth layer. Basically, this is the listening algorithm of Equation 3 with some Gaussian noise at relevant locations to create space (uncertainty) for reasoning about the user’s preferred settings. The unusual thing about Equation 6 is that the observation variance For the second model 34 , which, as we will see later, reflects the end-user’s assessment of the current hearing device behavior (related to confidence).

[0146] Adjusting parameters is also considered to be subject to a dynamical system of the form:

[0147]

[0148]

[0149] In this system, is the (hidden) state variable of the second layer, and Modeled as an "observation model" that depends on the upper state of the user model (first model 32) and second-tier assessment status This layered model of the hearing device signal processing system can be continued to higher layers. Figure 3 In , an N-layer hierarchical dynamical system is derived for the hearing device algorithm. (n=1, .., N), which are also considered as hidden states in the agent model 30. The agent 8 calculates the posterior distribution of these parameter states through (message passing) reasoning. At any time, the agent 8 can be based on the relative The agent trusts that a set of test values ​​of the parameter is submitted to the hearing device 6 exist Figure 3 , shows how to go from relative to The posterior distribution of Submit a set of parameter values ​​to the hearing device 6 is called an attempt (or action; by agent 8). A typical method is to start from the parameter This is often called Thompson sampling.

[0150] Users can rate or provide feedback on the current hearing device behavior at any time. t , i.e. providing evaluation data. Feedback (evaluation) from the user may include pressing a button to indicate like / dislike, or may include recording of physiological signals such as heart rate and EEG signals, or a combination thereof.

[0151] Agent 8 will these evaluations r t (also called evaluation data) is considered as another (set of) observation signals. The agent 8 includes a second model 34, which is a generated probabilistic user model for predicting these evaluations. In a similar manner to the hearing device model (first model 32), the user model (second model 34) is also a hierarchical probabilistic system. The mth layer in the user model is generally given by:

[0152]

[0153]

[0154] Wherein, Formula 8a describes the user model as having a hidden state (Markov) dynamical system. The transition probability between states is determined by the control variable Control. According to Equation 8b, these control variables are functions of the upper layer state. Figure 3 At the first layer in is also generated as a (probability) function of the state. In the second model, the hidden state Corresponding to the evaluation data r t The reason for the prediction.

[0155] It should be pointed out that in the future, ＞t =(r t+1 ,r t+2 , ...) are part of the second model 34. The second model 34 optionally allows setting the prior expectation p(r ＞t ),like Figure 3 The termination and r ＞t The corresponding edges are shown as nodes 40, 42. These prior expectations of future observations constrain the inferences in the model to take into account that future observations will be evaluated positively. By including these prior constraints, no matter what environment the second model is used in, any reasoning action will take into account that future evaluations are expected to be evaluated positively.

[0156] Optionally, the constraints of model 30 are not limited to future observations r ＞t Set constraints. You can set constraints such as u ＞t It is noted that the agent's model 30 includes links 36, 38 that directly connect variables in the hearing device signal processing model 32 with variables in the user model 34. For example, the link 36 connects the hidden user model state Connect to the nodes in the signal processing model 32 An exemplary node function ("observation model") would be In this equation, the variance of the observation model depends on the state of the user model 34. Therefore, in this model, the estimated "satisfaction state" affects the observation y t If a high user satisfaction state is associated with a high precision (i.e. a small variance) for the observed hearing device signal output (and vice versa, a low user satisfaction state is associated with a low precision for the hearing device output signal observation), the model 30 of the agent 8 will only learn from the HA behaviors associated with the high user satisfaction state. In this way, the user model provides a context for modulating the learning rate of the hearing device signal processing model.

[0157] The above set of equations provides a typical example of a model 30 of a probabilistic agent 8 that designs a listening algorithm for interaction with an end user 12. The details of an actual implementation may vary.

[0158] Reasoning and learning via message passing

[0159] The agent's model 30 specifies the observed variables x t ,y t ,r t , (unobserved) future observation r ＞t , intervention variables (action, trial) θ t and hidden (unobserved) state variables To simplify the notation for inference and learning, the variables are grouped into a set of observed variables d at time step t t ={x t ,y t ,r t} and unobserved variables Technically, the tuning parameters and hyperparameters of model 30 should be added to the set of unobserved variables (z t ). The agent 30 has only one goal: given the observed variables and given the model assumptions, make Bayesian inferences about the unobserved variables (including reasoning about their actions). This is usually a difficult problem. In the Forney-type factor graph framework, Bayesian inference can be approximated by a message passing algorithm. Examples of message passing algorithms for FFGs include belief propagation, variational message passing, and expectation propagation. Variational message passing is a message passing implementation for variational inference. In variational inference, a so-called free energy functional F is defined:

[0160]

[0161] This is a function of the variational distribution q(z). Minimization of the free energy F[q] with respect to q leads to approximate Bayesian inference. F can be minimized by message passing in the FFG. Thus, by applying message passing-based inference in the FFG of the probabilistic model, e.g. Figure 3 As shown, the agent 8 will design a hearing algorithm that optimizes the expected future user satisfaction rate, thereby providing an improved hearing algorithm.

[0162] The use of the terms "first", "second", "third" and "fourth", "primary", "secondary", "tertiary", etc. does not imply any particular order, but rather is included to identify individual elements. Furthermore, the use of the terms "first", "second", "third" and "fourth", "primary", "secondary", "tertiary", etc. does not indicate any order or importance, but rather "first", "second", "third" and "fourth", "primary", "secondary", "tertiary", etc. are used to distinguish one element from another. It should be noted that the use of the terms "first", "second", "third" and "fourth", "primary", "secondary", "tertiary", etc., here and elsewhere, is for labeling purposes only and is not intended to indicate any particular spatial or temporal order.

[0163] Furthermore, the marking of a first element does not imply the presence of a second element, and vice versa.

[0164] Understandably, Figure 1-3 Include some modules or operations shown with solid lines and some modules or operations shown with dotted lines. The modules or operations included in the solid lines are the modules or operations included in the most extensive example embodiments. The modules or operations included in the dotted lines are example embodiments of other modules or operations that can be included in other modules or operations or as a part of them or just adopted in addition to the modules or operations of the solid line example embodiments. It should be understood that these operations do not need to be performed in the order presented. In addition, it should be understood that not all operations need to be performed. The exemplary operations can be performed in any order and in any combination.

[0165] It should be noted that the word "comprising" does not necessarily exclude the presence of other elements or steps than those listed.

[0166] It shall be noted that the word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements.

[0167] It should also be noted that any reference signs do not limit the scope of the claims, that exemplary embodiments may be implemented at least partially by means of both hardware and software, and that several "means", "units" or "devices" may be represented by the same hardware.

[0168] Various exemplary methods, devices, agents, and systems described herein are described in the general context of method steps and processes, which can be implemented in one aspect by a computer program product including computer executable instructions (such as program code executed by a computer in a network environment) implemented in a computer-readable medium. Computer-readable media may include removable and non-removable storage devices, including but not limited to read-only memory (ROM), random access memory (RAM), compact disk (CD), digital versatile disk (DVD), etc. In general, program modules may include routines, programs, objects, components, data structures, etc. that perform specified tasks or implement specific abstract data types. Computer executable instructions, associated data structures, and program modules represent examples of program codes for performing the steps of the methods disclosed herein. A specific sequence of such executable instructions or associated data structures represents an example of corresponding actions for implementing the functions described in such steps or processes.

[0169] Although features have been shown and described, it should be understood that they are not intended to limit the claimed invention, and it will be apparent to those skilled in the art that various changes and modifications may be made without departing from the spirit and scope of the claimed invention. Therefore, the specification and drawings are to be regarded as illustrative rather than restrictive. The claimed invention is intended to cover all alternatives, modifications and equivalents.

Claims

1. A method for contextual design of a hearing algorithm of a hearing device, the method being performed in situ by an accessory device of a hearing system comprising the hearing device and an accessory device, the method comprising the following steps: Initializing the model including parameterizing the objective function; providing one or more operating parameters indicative of the hearing algorithm to the hearing device; Acquiring operation data including corresponding input data and output data of the hearing device, the step comprising: receiving said input data from said hearing device, and determining or calculating corresponding output data by applying a current hearing algorithm executed by the hearing device; obtaining evaluation data indicative of a user evaluation of the output data; determining one or more updated operating parameters based on the model, the operating data, and the evaluation data; and providing the updated operating parameters to the hearing device, wherein providing the one or more operating parameters to the hearing device comprises providing a hearing algorithm identifier to the hearing device, Wherein, the model is a generative probability model, and Wherein, the generation probability model is the generation probability model m given by the following formula: p(x,y,r,f,θ,z|m), Wherein, x is the input data of the hearing device, y is the output data of the hearing device, r is the evaluation data, f is the hearing algorithm, θ is the hearing device parameter, and z is the latent variable.

2. The method according to claim 1, wherein: Providing one or more operating parameters to the hearing device comprises providing one or more hearing device parameters to the hearing device.

3. The method according to claim 1, wherein: The model is a hierarchical Bayesian dynamics system.

4. The method according to any one of claims 1 to 3, wherein: The hearing algorithm is a dynamical system given by: s t+1 =f(s t ,x t ,θ) y t =g(s t ) Among them, t is the time mark; x t is the input data, s t is the state vector, y t As state s t , f(·) represents the hearing algorithm, and θ comprises a set of hearing device parameters.

5. The method according to any one of claims 1 to 3, wherein: Determining one or more updated operating parameters includes solving a probabilistic inference problem given by: p(f,θ|d <t ={(x,y,r) <t },r >t >0,m), Where f is the hearing algorithm, θ is the hearing device parameter, and d <t ={(x,y,r) <t } is the first constraint on past observations, and r >t >0 is a second constraint that future evaluation data should indicate positive user evaluations of the output data.

6. The method according to claim 5, wherein: Solving the probabilistic inference problem includes implementing message passing on a factor graph implementation of the model.

7. The method according to claim 6, wherein: The factor graph is implemented as a Forney-type factor graph implementation of the model.

8. The method according to any one of claims 1 to 7, wherein: Determining one or more updated operating parameters based on the model, the operating data, and the evaluation data includes minimizing a function, wherein the function is an alpha divergence function.

9. The method according to claim 8, wherein: The alpha divergence function is a function F given by: Among them, the function F is the function of the variational distribution q(z), z is the latent variable, and d is the data tuple.

10. The method according to any one of claims 1 to 9, wherein: The model includes a first model and a second model, wherein the first model is a generative probability model for the input data and the output data, and the second model is a generative probability model for the evaluation data, wherein the first model is based on an output of the second model.

11. An accessory device for a hearing system, the accessory device comprising a memory module, a processor module and a wireless interface, wherein: The accessory device is configured to perform the method according to any one of claims 1 to 10.

12. A hearing system comprising a hearing device and an accessory device according to claim 11.

Images