Modeling the Hearable Case
The machine learning technology automates the shell design of ear-mounted devices, which solves the problem of time-consuming, expensive and inconsistent manual design in the prior art, and achieves a fast, economical and high-quality customized ear-mounted device shell design.
Patent Information
- Application Number
- CN202211063557.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2018-11-16
- Filing Date
- 2019-11-15
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2039-11-15
AI Technical Summary
The housing design of existing ear-mounted devices relies on manual modeling of highly skilled operators, is time-consuming and inconsistent, making it difficult to meet the needs of fast customization and cost-effectiveness.
Using machine learning technology, based on data from previously manufactured ear-mounted devices and user ear impressions, the training model generates an automated ear-mounted device housing design that mimics the manual design of highly skilled operators.
Reduces manual intervention in design and modeling, reduces cost and inconsistency, improves productivity and product quality, and enables the generation of customized ear-mounted device housings that meet user needs in minutes.
Smart Images

Figure CN115426610B_ABST
Abstract
Description
[0001] This application is a divisional application of the Chinese national phase application of the PCT application with an application date of November 15, 2019, an international application number of PCT / US2019 / 061807, and an invention name of “Modeling of Ear-worn Device Housing”. The entry date of this Chinese national phase application into the Chinese national phase is May 14, 2021, and the application number is 201980075394.8.
[0002] Cross-reference to related applications
[0003] This application claims the benefit of U.S. Provisional Patent Application No. 62 / 768,208, filed on November 16, 2018, the entire contents of which are incorporated herein by reference. Technical Field
[0004] The present disclosure relates to ear-worn devices. Background Art
[0005] An ear-worn device is a device designed to be worn on or in a user's ear. Example types of ear-worn devices include hearing aids, headphones, earbuds, telephone handsets, and other types of devices designed to be worn on or in a user's ear.
[0006] Some manufacturers of hearables rely on highly skilled operators to design hearables using 3D modeling software. When the hearables are produced, these highly skilled operators and / or audiologists may make physical modifications to the housing of the hearables, e.g., to alleviate discomfort to the user, or to shape the housing based on the needs of the manufacturer or clinician. Manually modeling and shaping hearables in this manner is time consuming, expensive, and can result in inconsistencies due to, for example, variations in operator skill levels and techniques. Summary of the invention
[0007] In general, this disclosure describes techniques for enabling a computing device to use machine learning techniques to automatically model an ear-worn device or housing having properties similar to an ear-worn device manually created by a highly skilled operator based on an original ear impression of a target ear canal.
[0008] In one example, a method is described that includes obtaining, by at least one processor, a representation of a target ear canal of a user. The method also includes generating a representation of an ear-worn device of the target ear canal using a machine learning model trained at least in part based on representations of previously manufactured ear-worn devices.
[0009] In another example, a computing device is described that includes at least one processor configured to obtain a representation of a target ear canal of a user. The at least one processor is further configured to generate an ear-worn device representation of the target ear canal using a machine learning model trained at least in part based on representations of previously manufactured ear-worn devices.
[0010] In another example, a computer-readable storage medium comprising instructions is described that, when executed by at least one processor, configure the at least one processor to obtain a representation of a target ear canal of a user. When executed, the instructions further configure the at least one processor to generate a representation of an ear-worn device for the target ear canal using a machine learning model trained at least in part based on representations of previously manufactured ear-worn devices.
[0011] In another example, a system is described that includes means for obtaining a representation of a target ear canal of a user. The system also includes means for generating an ear-worn device representation of the target ear canal using a machine learning model trained at least in part based on representations of previously manufactured ear-worn devices.
[0012] The details of one or more aspects of the disclosure are set forth in the accompanying drawings and the following description. Other features, objects, and advantages of the techniques described in the disclosure will be apparent from the specification, drawings, and claims. BRIEF DESCRIPTION OF THE DRAWINGS
[0013] Figure 1 is a conceptual diagram illustrating an example machine learning system according to one or more aspects of the present disclosure.
[0014] Figure 2 is a block diagram illustrating an example computing device configured to train and execute a machine learning model according to one or more aspects of the present disclosure.
[0015] Figure 3A is a block diagram illustrating an example ear-worn device modeled by an example machine learning model according to one or more aspects of the present disclosure.
[0016] Figure 3B is a conceptual diagram illustrating a cross-sectional view of a representation of a target ear canal including a position along an axis protruding through the target ear canal according to one or more aspects of the present disclosure.
[0017] Figure 4 is a flow diagram illustrating operations performed by an example machine learning system according to one or more aspects of the present disclosure. DETAILED DESCRIPTION
[0018] An ear-worn device is a device designed to be worn on or in the ear of a user. Example types of ear-worn devices include hearing aids, headphones, earbuds, telephone handsets, and other types of devices designed to be worn on or in the ear of a user. As used herein, ear-worn devices such as hearing aids, hearing devices, and hearing instruments refer to any ear-worn device used as a hearing aid, personal sound amplification product (PSAP), headset, hearable, wired or wireless earbud, or other hearing instrument that provides sound to a user for hearing.
[0019] Some hearable professionals take ear impressions of patients or end users and send the original impressions and / or scans of the original impressions along with other patient-specific requirements (e.g., style, features, etc.) to hearable manufacturing facilities. Highly skilled operators (e.g., product modelers, audiologists, etc.) at the manufacturing facility can scan the original impressions or import the scanned impressions into 3D modeling software.
[0020] Based on the patient's specific requirements and a set of empirical modeling guidelines known to the operator, the operator can manually design the ear-worn device, for example, by interacting with modeling software to create a computer model of a housing that can contain all of the internal components of the ear-worn device (e.g., microphones, receivers, circuits, vents, etc.) and fit comfortably in the ear of a specific user. Even with the help of three-dimensional modeling software, a skilled operator may spend 5 to 15 minutes manually creating a housing model. Once the operator is satisfied with the housing design, the ear-worn device is manufactured according to the housing design. Thereafter, the operator at the manufacturing plant and / or the ear-worn device professional at the clinic or retail location can spend additional time further modifying the housing of the ear-worn device (e.g., using hand tools) to minimize any actual or foreseeable end-user discomfort.
[0021] Manually modeling and shaping ear-worn devices in this manner is time consuming and expensive, and can result in inconsistencies due to, for example, differences in operator skill levels and operator techniques. If an end-user or professional wishes to preview their customized, ear-wearable design prior to manufacturing, the clinic or retailer must have access to a dedicated modeler, expensive modeling software, and at least 5 to 15 minutes of modeling time. Furthermore, for customized ear-worn devices, such as over-the-counter hearing aids, etc., the cost and time of relying on highly skilled and / or specialized operators to manually create ear-wearable designs may be unfeasible from a business perspective, and may introduce delays in the procurement of ear-worn devices that may be unacceptable to retailers, professionals, or end-users.
[0022] In general, the present disclosure describes techniques for enabling a computing device to use machine learning techniques to automatically model ear-worn devices with properties similar to finished ear-worn devices manually created by highly skilled operators and / or clinicians. For example, over time, some ear-worn device manufacturers and clinicians have accumulated hundreds of thousands (if not millions) of original ear impressions and corresponding data related to finished ear-worn device products. This data can span a variety of ear-worn device styles and other device and / or end-user characteristics. Instead of manually creating ear-worn devices using highly skilled operators and time-consuming modeling software, a machine learning model can be trained using the accumulated ear impressions and corresponding data associated with finished ear-worn devices to output ear-worn device designs that mimic finished ear-worn devices created by skilled modelers and clinicians.
[0023] Figure 1 is a conceptual diagram illustrating an example machine learning system according to one or more aspects of the present disclosure. Figure 1 The system 100 is configured to automatically model ear-hearable devices that have properties similar to finished ear-hearable devices that are manually created by highly skilled operators working in a manufacturing plant, clinic, or other retail store.
[0024] System 100 can be executed on one or more processors of a computing device or computing system. For example, system 100 can be executed on at least one processor of a computer workstation located at a manufacturing plant of ear-worn devices. In other examples, system 100 can be executed on at least one processor of a computer workstation located at a clinic or other retail store where ear-worn devices are installed and sold to patients and other end users. In some cases, system 100 can be executed in a cloud computing environment and can be remotely accessed via a computer workstation located at a manufacturing plant, clinic or other retail store. In other examples, system 100 can be executed as part of an application executed on at least one processor of a mobile computing device associated with a patient or other end user.
[0025] The system 100 includes a machine learning (ML) model 102. The ML model 102 is configured to generate a representation of an ear-worn device for a target ear canal of a user as output data 108. The ML model 102 is configured to generate a representation of an ear-worn device that simulates a finished ear-worn device produced by a highly skilled operator in a manufacturing plant, clinic, or other retail store. That is, the ML model 102 is trained to produce a representation of an ear-worn device whose shape and configuration are as close as possible to the shape and configuration of an ear-worn device that has been physically modified by a highly skilled operator in a manufacturing plant, clinic, or retail store.
[0026] ML model 102 may be or may include one or more of a variety of different types of machine learning models. For example, ML model 102 may include one or more neural networks, deep neural networks, parametric models, or other types of machine learning models. Figure 2 Other examples of such different types of machine learning models are provided in the description of machine learning (ML) model 202.
[0027] The ML model 102 is trained using training data 104. There are many examples of training data 104. As one example, the training data 104 can include representations of previously manufactured ear-worn devices and representations of corresponding ear impressions used to manufacture the previously manufactured ear-worn devices.
[0028] For example, over time, a clinician or a salesperson at a retail store may have obtained hundreds, if not thousands, of 3D impressions of target ear canals for end users and sent these ear impressions or scans thereof to a hearable manufacturer. After a highly skilled operator imports the ear impressions into 3D modeling software and uses hand tools to create the finished hearable, the clinician, retailer, or manufacturer may have recorded the dimensions and general characteristics of the finished hearable that is ultimately worn by the end user.
[0029] The training data 104 may include representations of ear impressions obtained over time and representations of finished wearable ear devices corresponding to the ear impressions. In other words, the training data 104 may include representations of actual ear impressions obtained over time as example input data, and may include representations of finished ear-worn devices corresponding to the example input data as desired output data. In this way, when the ML model 102 generates a representation of an ear-worn device for a particular target ear canal, the ML model 102 may infer design rules and constraints to be followed from the training data 104 so that the shape and design of the representation generated by the ML model 102 mimics the shape and design of an ear-worn device that would be manually generated by a skilled operator. In some examples, the training data 104 includes other data in addition to the historical ear impressions and corresponding final ear-worn device features, such as information about the training data (e.g., Figure 2 The training data in the training data memory 204 shown in FIG. 2 is described in more detail.
[0030] The ML model 102 is configured to receive input data 106, and in response, the ML model 102 is configured to provide output data 108. The input data 106 includes information specifying characteristics of a target ear canal and / or an end user. The output data 108 includes a representation of an ear-worn device that the ML model 102 determines more closely matches the shape and design of a finished ear-worn device that is manually created by a skilled operator using modeling software and / or other tools. For example, the output data 108 may include a three-dimensional representation of an ear-worn device housing that more closely conforms to areas of the target ear canal having a greater curvature than other areas of the target ear canal having a lesser curvature. The three-dimensional representation may include positioning information indicating the location of internal components within the housing.
[0031] Examples of input data 106 and output data 108 are provided below. Additional examples of input data 106 and output data 108 are also possible, such as with respect to input data (e.g., Figure 2 206) and output data (e.g., Figure 2 The output data in the output data memory 208 shown in FIG. 2 is described in more detail.
[0032] The input data 106 includes a representation of a target ear canal of an end user. The representation of the target ear canal can be a three-dimensional representation of an ear impression of the end user. In other examples, the representation of the target ear canal can be a three-dimensional representation based on an image (e.g., a photograph, x-ray, or other type of image) taken of the target ear canal of the end user. In some examples, the representation of the target ear canal can be refined or preprocessed in one or more ways before being passed as input to the ML model 102. For example, the representation of the target ear canal can be generated using a Fourier transform or other operation of a three-dimensional representation of an ear impression or image taken of the end user.
[0033] In some examples, the input data 106 includes an indication of a particular type of ear-worn device. For example, depending on the type of ear-worn device being generated, the ML model 102 can follow different design rules and constraints. In other words, the model 102 can be trained using training data 104 associated with a first type of ear-worn device, and the model 102 can be further trained using training data 104 associated with a second, different type of ear-worn device. When the input data 106 specifies a first type, the model 102 can generate a representation of the first type of ear-worn device, and when the input data 106 specifies a second, different type, the model 102 can generate a second, different representation of the second type of ear-worn device.
[0034] Also referred to as ear-worn styles, there are many examples of ear-worn types. As some examples, the indication of a particular type of ear-worn may specify whether the ML model 102 should generate: a hearing aid, a headset, an earplug, a telephone handset, or another type of ear-worn. With respect to hearing aids, or so-called hearing aids, hearing devices, or hearing instruments, the indication of a particular type of ear-worn may specify whether the ML model 102 should generate one of three types of hearing aids.
[0035] One type of hearing aid device includes a housing or shell that is designed to be worn on the ear for aesthetic and functional reasons. The housing or shell encloses the electronics of the hearing instrument. Such devices may be referred to as in-the-ear (ITE), in-the-canal (ITC), completely in-the-canal (CIC), or invisible-in-the-canal (IIC) hearing instruments. The ML model 102 may be configured to generate a design for a housing or shell that is designed to be worn in the ear as a representation of an ear-worn device for a target ear canal of an end user.
[0036] The input data 106 may also include a corresponding representation of each internal component of the ear-worn device designed by the ML model 102. The corresponding representation of each internal component of the ear-worn device may include a three-dimensional representation (e.g., physical dimensions) of the internal component or a Fourier transform or other refinement or manipulation of the three-dimensional representation of the internal component.
[0037] For example, an ear-worn device may include several internal components of different sizes. The final size or shape of the housing of the ear-worn device may depend on the size and shape of each internal component of the ear-worn device. ML model 102 may generate a representation of the ear-worn device that may accommodate the size and shape of each internal component selected to be included in the final design. Figure 3A An example of the internal components of an ear-worn device is described in more detail with respect to ear-worn device 300 .
[0038] The input data 106 may also include an indication of one or more modeling constraints to be followed by the ML model 102. The modeling constraints may include quantitative parameters derived from empirical modeling guidelines obtained from expert modelers and / or audiologists. In other words, the modeling constraints of the input data 106 may specify specific "rules of thumb" derived from skilled operator experience that are converted into quantitative parameters. The ML model 102 may be fed one or more modeling constraints to increase the likelihood that the ML model 102 generates a representation of an ear-worn device that fits the end user's ear with little or no physical modification.
[0039] As an example of a modeling constraint, the input data 106 can include one or more constraints on the relative positioning between two or more internal components or on the position of one or more internal components relative to the housing of the ear-worn device. For example, a minimum spacing between the faceplate, the receiver, and / or the hearing housing may be required to ensure operational stability. Similarly, a receiver may function properly unless placed at a minimum distance from the housing of the ear-worn device or otherwise prevented from contacting the housing.
[0040] As another example, input data 106 may include one or more constraints on the location of one or more internal components relative to an access panel of the ear-worn device. For example, a processor or circuit may need to be located near a faceplate or access panel of the ear-worn device.
[0041] As another example, input data 106 can include one or more constraints on the thickness of the housing of the ear-worn device. For example, while it may be desirable to reduce the housing wall thickness, e.g., to create a smaller and lighter-weight ear-worn device, a minimum thickness may be required to improve the integrity and durability of the device.
[0042] As another example, input data 106 may specify features of an ear-worn device. For example, an ear-worn device may include a vent for receiving sound waves and / or outputting sound waves. Over the years, skilled operators may have recognized that a particular size of the vent is critical to ensuring good sound quality. Input data 106 may include information specifying a minimum cross-sectional area or other size of such a vent.
[0043] The ML model 102 is configured to receive feedback 110, in the form of training data 104, for retraining the ML model 102 based on modifications to the output data 108 and the representation of the ear-worn device for the target ear canal. In other words, the representation of the ear-worn device generated by the ML model 102 may initially include errors or require physical modifications to improve comfort for the end user. However, over time, by providing feedback 110 about the errors to the ML model 102, the ML model 102 can continuously learn and improve, so that subsequent representations of the ear-worn device generated by the ML model 102 require little or no physical modifications. Thus, in some examples, the ML model 102 can be retrained based on the user's modifications to the representation of the ear-worn device for the target ear canal.
[0044] As one example, feedback 110 may include a specific dimensional change made by a highly skilled operator to a housing of an ear-worn device to better fit the ear-worn device within a target ear canal of a user. The dimensional change may include an actual dimensional change or a refinement or manipulation of the dimensional change. As another example, feedback 110 may include a specific positional change made by a highly skilled operator to the positioning of an internal component within a housing of an ear-worn device (e.g., to comply with empirical design guidelines). The positional change may include an actual positional change or a refinement or manipulation of the positional change.
[0045] In some examples, the training data 104 includes other data in addition to the historical ear impressions and the corresponding final ear-worn device characteristics. For example, the training data 104 may also include characteristics of the corresponding end users of the final, previously manufactured ear-worn devices. For example, over time, the same clinician or retail store salesperson may have taken hundreds (if not thousands) of three-dimensional impressions of the target ear canals for the end users, and may have recorded other information about the end users. That is, while taking precautions to protect or ignore any personally identifiable information in the user data collected over time, the clinician or retail store salesperson may retain the characteristics of the end user, including each of the end user's age, gender, nationality, geographic location, or other personal characteristics. The clinician or retail store salesperson may retain specific information about the end user's ears, including information specifying a specific ear type, a specific ear shape, and / or a specific ear or ear canal size.
[0046] When the ML model 102 generates a representation of an ear-worn device for a particular target ear canal, the ML model 102 can infer additional design rules and constraints to follow from the characteristics of these users. For example, following a rule or constraint can cause the model 102 to generate a representation of an ear-worn device whose shape and design mimic the shape and design of an ear-worn device manually produced by a skilled operator. For example, the ML model 102 can infer that for a particular end user's age, gender, nationality, geographic location, or other personal characteristic, an ear-worn device manually produced by a skilled operator has a particular shape or design that is unique to that personal characteristic. In some examples, the ML model 102 may include multiple models, each trained to handle a different age, gender, nationality, geographic location, or other personal characteristic.
[0047] In some examples, training data 104 includes characteristics or preferences of certain clinics, retailers, or user groups. For example, over time, certain clinicians or salespeople may have developed their own "brand" of devices, adhering to their own specific preferences when designing custom ear-worn devices. When ML model 102 generates representations of ear-worn devices for a particular clinic or clinician, ML model 102 can be trained to learn these clinician preferences and infer additional design rules and constraints to follow.
[0048] System 100 can reduce the amount of human intervention required to design and model hearables by executing ML model 102, thereby reducing costs and improving overall quality by reducing inconsistencies that may be observed between different modelers. Executing ML model 102 can not only save manpower and modeler training, but can also reduce the overall housing modeling and modification time by at least several times, thereby potentially saving costs.
[0049] Figure 2 is a block diagram illustrating an example computing device configured to train and execute a machine learning model according to one or more aspects of the present disclosure. Figure 2 The computing device 212 is configured to train and / or execute a machine learning (ML) model 202, which is Figure 1 An example of a machine learning model 202. Figure 2 Only one particular example of computing device 212 is shown, and many other example configurations of computing device 212 exist.
[0050] The computing device 212 may include one or more electronic computing devices, including a combination of one or more of the following: a mobile device, a wearable device (e.g., ear wearable or non-ear wearable), a server device, and other types of computing devices. Some specific examples of the computing device 212 include a combination of one or more of the following: a mobile phone, a tablet computer, a portable media player device, a stand-alone server device, a blade server, a personal computer, a workstation, a content delivery network device, and another type of device. Therefore, in general, the description of the computing device 212 performing a specific action or function in this disclosure should be interpreted as a combination of one or more electronic computing devices performing a specific action or function.
[0051] like Figure 2 , computing device 212 includes one or more processors 222, one or more input components 218, one or more communication units 214, one or more output components 220, data storage 216 (e.g., one or more data storage devices), and communication bus 224. Computing device 212 may include many other components.
[0052] The communication bus 224 can interconnect (physically, communicatively, and / or operatively) each of the components 214, 216, 218, 220, and 222 for inter-component communication. In some examples, the communication bus 224 can include a system bus, a network connection, an inter-process communication data structure, or any other method for transmitting data.
[0053] The computing device 212 may rely on input components 218 to receive user input. Examples of user input include tactile, auditory, and visual input. The input components 218 may include a presence-sensitive screen, a touch-sensitive screen, a camera, a mouse, a track pad, a trackball, a keyboard, a microphone, a physical button, a communication port, a sensor, or any other type of component configured to detect input from a person or a machine.
[0054] The computing device 212 may rely on output components 220 to generate output. Examples of output include tactile, auditory, and visual output. The output components 220 may include a display, such as a presence-sensitive screen, a liquid crystal display (LCD), or other type of display, a sound card, a video graphics adapter card, a speaker, or any other type of component configured to produce output to a person or machine. In some examples, a single component may span input components 218 and output components 220, for example, a presence-sensitive screen is an example of an input component 218 and an output component 220.
[0055] The communication unit 214 may enable the computing device 212 to exchange data with one or more other computing devices (e.g., via a communication network, such as a local area network or the Internet). The communication unit 214 may include a transmitter and a receiver that enable the computing device 212 to communicate with other computing devices, either wired or wirelessly. Examples of the communication unit 214 may include an optical transceiver, a radio frequency transceiver, or other types of devices capable of sending and receiving information. Other examples of the communication unit 214 may include a Bluetooth radio, a cellular data radio, a Wi-Fi radio, a Thunderbolt or Universal Serial Bus (USB) interface, etc.
[0056] Communication unit 214 can configure computing device 212 to communicate with a remote computing device that maintains a database or other record of previously manufactured ear-worn devices. For example, computing device 212 can use communication unit 214 to receive data from a production facility database, including representations of previously manufactured ear-worn devices for training a machine learning model (e.g., ML model 202). As another example, computing device 212 can use communication unit 214 to receive other data from a clinician or retail store database, including representations of previously manufactured ear-worn devices for training a machine learning model (e.g., ML model 202).
[0057] The computing device 212 includes one or more processors 222, which are configured to read instructions from the data storage 216 and execute the instructions to cause the computing device 212 to take actions or perform functions. For example, according to one or more aspects of the present disclosure, the processor 222 can execute instructions that cause the computing device 212 to train and execute a machine learning model (e.g., ML model 202).
[0058] Data storage 216 may store information needed during operational use of computing device 212, including instructions executed by processor 222. In some examples, the primary purpose of data storage 216 is to serve as a short-term rather than long-term computer-readable storage medium. Data storage 216 may be volatile memory, so if computing device 212 loses power, data storage 216 does not retain stored contents. Data storage 216 may also be configured to store information long-term as a non-volatile memory space and retain the information even if computing device 212 loses power.
[0059] like Figure 2 As shown, data storage 216 includes information associated with ML model 202, training module 226, and user interface (UI) module 228. In addition, data storage 216 includes input data storage 206, training data storage 204, and output data storage 208. Processor 222 can execute instructions associated with ML model 202 and modules 226 and 228 to cause computing device 212 to perform functions.
[0060] The UI module 228 is configured to manage a user interface associated with the computing device 212. Specifically, according to the techniques of this disclosure, the UI module 228 can provide an interface (e.g., a graphical user interface) from which a user of the computing device 212 can interact with the ML model 202 to cause the ML model 202 to generate a representation of the ear-worn device based at least in part on the representation of the target ear canal. For example, the UI module 228 can cause the display screen of the output component 220 to provide a graphical or other type of user interface from which the user can view the representation of the ear-worn device generated by the ML model 202. When the user interacts with the user interface, the UI module 228 can interpret the input detected by the input component 218, for example, to provide user-specific controls to the ML model 202. The UI module 228 can exchange information with the ML model 202 and / or the training module 226.
[0061] As one example, computing device 212 can be a production workstation executing production modeling software that includes ML model 202 and modules 226 and 228. UI module 228 can interpret inputs and generate outputs associated with the production modeling software, e.g., to enable a user to control various parameters of ML model 202. For example, a user of computing device 212 can provide inputs detected by input component 218 that cause UI module 228 to perform functions, e.g., configure ML model 202 to generate a representation of a particular type of hearable with particular components selected by the user. A skilled operator at a production facility can make adjustments to ML model 202, e.g., based on discussions or information obtained from a clinician or quality control specialist at final assembly or in the field.
[0062] As another example, computing device 212 can be a clinician workstation executing clinician modeling software, similar to the production modeling software mentioned above, but focused on addressing the needs of clinicians, audiologists, or retailers rather than the needs of production operators. For example, a user of computing device 212 can provide input detected by input component 218, which causes UI module 228 to perform functions, such as configuring model 202 to generate a representation of an ear-worn device that addresses specific requirements indicated to the user by a patient during a clinic consultation. Thus, professionals in a clinic are able to interact with computing device 212 to quickly model and visualize different types of ear-worn devices on behalf of a patient or client, enabling professionals and clients to make informed decisions before sending a request to a manufacturer to produce an ear-worn device.
[0063] As another example, computing device 212 may be a mobile device, such as a tablet or mobile phone, that executes a consumer application that enables a user to interact with consumer modeling software that includes (or at least provides access via the cloud to) ML model 202 and modules 226 and 228. UI module 228 may interpret input and generate output associated with the consumer application, e.g., to enable an end user to exercise some control over various parameters of ML model 202. For example, a user of computing device 212 may provide input detected by input component 218 that causes UI module 228 to perform functions, such as configuring ML model 202 to generate a representation of an ear-worn device that satisfies general usability characteristics and preferences that have been selected by the end user through the consumer application. The ear-worn device user himself may be able to quickly visualize different types of customized ear-worn device products without going to a clinic and make an informed decision to order the product at home or on-site. Thus, once a patient or other end user submits an ear impression or scan to a consumer application (e.g., performed at a pharmacy), the user can obtain and visualize a simulated ear-worn shell model in just minutes or seconds and make a quick decision about ordering the ear-worn device on the spot, and can be assured that the resulting ear-worn device will sound great, look good, and be comfortable to wear.
[0064] UI module 228 can write information to input data storage 206 and can read information from output data storage 208. UI module 228 can further write feedback information to training data storage 204, e.g., the feedback information can indicate any changes made by the user to the final ear-worn device design generated by ML model 202.
[0065] The training module 226 trains the ML model 202 based on the information contained in the training data storage 204. The training module 226 can be configured to continuously train the ML model 202, for example, in response to receiving feedback information from the UI module 228 or by determining errors in the output of the ML model 202. The training module 226 is configured to convert the ML model 202 into a training mode, wherein the ML model 202 receives example input data and example output data so that the ML model 202 creates or updates any inferences previously established by the ML model 202 during a previous training session. The training module 226 can be configured to train the ML model 202 based on records of previously manufactured ear-worn devices, for example, records received from a manufacturing facility database, a clinician or retail store database, or other repository using the communication unit 214.
[0066] In some examples, training module 226 can determine errors or deviations between the expected output used to train ML model 202 and the output actually obtained from ML model 202. Training module 226 can emphasize fixing some errors more than others. For example, training module 226 can apply higher weights to certain areas of the representation of the ear-worn device to increase the likelihood that ML model 202 reduces errors in those areas, and apply lower weights to other areas to reduce the likelihood that ML model 202 reduces errors in those other areas.
[0067] For example, the representation of the ear-worn device may include a set of three-dimensional points that define the shape of the ear-worn device. Some three-dimensional points may be more important than other points (e.g., to ensure user comfort when wearing the ear-worn device). Training module 226 can determine the deviation of the shape of the ear-worn device from the desired shape and apply weights to the more important three-dimensional points where the deviation occurs to make ML model 202 more closely match the desired shape, especially at the more important points.
[0068] ML Model 202 is Figure 1 2. Like ML model 102, ML model 202 is configured to generate a representation of an ear-worn device for a target ear canal of a user at output data storage 208. ML model 202 is configured to generate a representation or ear-worn device that simulates a finished ear-worn device produced by a highly skilled operator in a manufacturing plant, clinic, or other retail store. That is, ML model 202 is trained to produce a representation of an ear-worn device whose shape and configuration are as close as possible to the shape and configuration of an ear-worn device that has been physically modified by a highly skilled operator in a manufacturing plant, clinic, or retail store.
[0069] The training module 226 trains the ML model 202 to receive input data from the input data store 206 and, in response, provide output data at the output data store 208. The input data may include input data of different types, forms, or variations.
[0070] During training, the training module 226 can intentionally deform the example input data or expected output data stored in the training data storage 204 in any number of ways to increase model robustness, generalization, or other qualities. Example techniques for deforming training data include adding noise; changing color, shading, or hue; amplification; segmentation; enlargement, etc.
[0071] The training module 226 can train the ML model 202 in an offline manner or in an online manner. In offline training (also known as batch learning), the training module 226 trains the ML model 202 on the entire static training data set. In online learning, the training module 226 continuously trains (or retrains) the ML model 202 as new training data becomes available (e.g., based on feedback from users).
[0072] The training module 226 can train the ML model 202 according to one or more different training types or techniques. For example, in some implementations, the training module 226 can train the ML model 202 using supervised learning, wherein the ML model 202 is trained on a training data set that includes instances or examples with labels. The labels can be applied manually by experts, generated through crowdsourcing, or provided by other techniques (e.g., physics-based or complex mathematical models). In some implementations, this process can be referred to as personalized model.
[0073] The training module 226 can train the ML model 202 by optimizing an objective function. For example, in some implementations, the objective function can be or include a loss function that compares (e.g., determines the difference between) output data generated by the ML model 202 from the training data and labels (e.g., ground truth labels) associated with the training data. For example, the loss function can calculate the sum or average of the squared differences between the output data and the labels. In some examples, the objective function can be or include a cost function that describes the cost of a particular result or output data. Other examples of objective functions can include edge-based techniques, such as triplet loss or maximum edge training.
[0074] The training module 226 can perform various optimization techniques to optimize the objective function. For example, the optimization technique can minimize or maximize the objective function. Example optimization techniques include Hessian-based techniques and gradient-based techniques, such as coordinate descent; gradient descent (e.g., stochastic gradient descent); sub-gradient methods, etc. Other optimization techniques include black-box optimization techniques and heuristic techniques.
[0075] The training module 226 can perform back propagation of errors to the ML model 202, which can be used in conjunction with optimization techniques (e.g., gradient-based techniques) to train the ML model 202 (e.g., when the machine learning model 202 is a multi-layer model such as an artificial neural network). For example, an iterative cycle of propagation and model parameter (e.g., weight) updates can be performed to train the ML model 202. Example back propagation techniques include truncated back propagation through time, Levenberg-Marquardt back propagation, and the like.
[0076] The training module 226 can use unsupervised learning techniques to train the ML model 202. Unsupervised learning can include inferring functions that describe hidden structures from unlabeled data. For example, classification or categorization may not be included in the data.
[0077] The training module 226 can train the ML model 202 using semi-supervised techniques that combine aspects of supervised learning and unsupervised learning. The training module 226 can train the ML model 202 through evolutionary techniques or genetic algorithms. The training module 226 can train the ML model 202 using reinforcement learning. In reinforcement learning, an agent (e.g., a model) can take actions in an environment and learn to maximize rewards and / or minimize penalties resulting from those actions. Reinforcement learning differs from supervised learning problems in that no correct input / output pairs are given and no suboptimal behavior is explicitly corrected.
[0078] The training module 226 can train the ML model 202 using one or more generalization techniques to improve the generalization of the ML model 202. The generalization techniques can prevent the ML model 202 from overfitting the training data. Examples of generalization techniques include dropout techniques; weight decay techniques; batch normalization; early stopping; subset selection; stepwise selection, etc.
[0079] ML model 202 can be or include one or more different types of machine learning models. One or more of the example models described below can be used (e.g., combined) to provide output data in response to input data. Additional models other than the example models provided below can also be used.
[0080] In some implementations, the ML model 202 can be a parametric model, while in other implementations, the ML model 202 can be a non-parametric model. For example, as a parameter-based model, the ML model 202 can determine a parametric representation of the ear-worn device divided into a plurality of cross-sections. The ML model 202 can determine a set of parameters that characterize a cylindrical shape that defines the ear-worn device, and then estimate these parameters for a given representation of a target ear canal. However, as a non-parametric model, such as a neural network, the ML model 202 determines an inherent structure associated with the ear-worn device and predicts the ear-worn device structure for a given target ear canal.
[0081] In some examples, the ML model 202 may be a linear model, while in other implementations, the ML model 202 may be a nonlinear model. In some implementations, the ML model 202 may be or may include one or more classifier models, such as a linear classification model, a quadratic classification model, etc. The ML model 202 may be or may include one or more regression models, such as a simple linear regression model, a multivariate linear regression model, a logistic regression model, a stepwise regression model, a multivariate adaptive regression spline, a local estimation scatter plot smoothing model, etc. In some examples, the ML model 202 may be or may include one or more decision tree-based models, such as a classification and / or regression tree, an iterative bisection 3 decision tree, a C4.5 decision tree, a chi-square automatic interaction detection decision tree, a decision tree stump, a conditional decision tree, etc.
[0082] The ML model 202 may include one or more instance-based learning models, such as a learning vector quantization model, a self-organizing map model, a local weighted learning model, etc. In some implementations, the ML model 202 may be or may include one or more nearest neighbor models, such as a k-nearest neighbor classification model; a k-nearest neighbor regression model, etc. The ML model 202 may be or include one or more Bayesian models, such as a naive Bayesian model, a Gaussian naive Bayesian model; a multinomial naive Bayesian model, an average single correlation estimator, a Bayesian network; a Bayesian belief network, a hidden Markov model, etc.
[0083] In some examples, ML model 202 may include one or more artificial neural networks (also referred to as neural networks). A neural network may include a set of connected nodes, which may also be referred to as neurons or perceptrons. A neural network may be organized into one or more layers. A neural network that includes multiple layers may be referred to as a "deep" network. A deep network may include an input layer, an output layer, and one or more hidden layers located between the input layer and the output layer. The nodes of a neural network may be connected or incompletely connected.
[0084] The ML model 202 may be or include one or more feed-forward neural networks. In a feed-forward network, the connections between nodes do not form a loop. For example, each connection may connect a node in a previous layer to a node in a subsequent layer.
[0085] In some examples, the ML model 202 may include one or more recurrent neural networks. In some cases, at least some nodes of the recurrent neural network may form a loop. Example recurrent neural networks include long short-term (LSTM) recurrent neural networks; gated recurrent units; bidirectional recurrent neural networks; continuous-time recurrent neural networks; neural history compressors; echo state networks; Elman networks; Jordan networks; recurrent neural networks; Hopfield networks; fully recurrent networks; sequence-to-sequence configurations, etc.
[0086] The ML model 202 may be or may include one or more convolutional neural networks. In some cases, the convolutional neural network may include one or more convolutional layers that perform convolution on the input data using a learned filter. The filter may also be referred to as a kernel.
[0087] In some examples, ML model 202 can be or include one or more generative networks, such as a generative adversarial network. The generative network can be used to generate new data, such as new images or other content.
[0088] The ML model 202 may be or may include one or more other forms of artificial neural networks, such as a deep Boltzmann machine; a deep belief network; a stacked autoencoder, etc. Any of the neural networks described herein may be combined (e.g., stacked) to form a more complex network.
[0089] The ML model 202 may include one or more clustering models, such as a k-means clustering model; a k-medians clustering model; an expectation-maximization model; a hierarchical clustering model, and the like.
[0090] The ML model 202 may receive and use the input data in its raw form, while in other cases, the raw input data may be preprocessed. For example, preprocessing the input data may include extracting one or more additional features from the raw input data.
[0091] Feature extraction techniques may be applied to the input data to generate one or more new additional features. Example feature extraction techniques include edge detection; corner detection; blob detection; ridge detection; scale-invariant feature transform; motion detection; optical flow; Hough transform, etc. The extracted features may include or be derived from the transformation of the input data to other domains and / or dimensions. As an example, the extracted features may include or be derived from the transformation of the input data to the frequency domain. For example, a wavelet transform and / or Fourier transform may be performed on the input data to generate additional features. In some examples, the extracted features may include statistics calculated based on the input data or certain portions or dimensions of the input data. Example statistics include the mode, mean, maximum, minimum, or other metric of the input data or portions thereof.
[0092] As another example of a preprocessing technique, part of the input data may be estimated. For example, additional synthetic input data may be generated by interpolation and / or extrapolation.
[0093] As another example preprocessing technique, some or all of the input data may be scaled, standardized, normalized, generalized, and / or regularized. Examples of regularization techniques include ridge regression; least absolute shrinkage and selection operator (LASSO); elastic net; least angle regression; cross validation; L1 regularization; L2 regularization, etc. As an example, some or all of the input data may be normalized by subtracting the mean of the eigenvalues for a given dimension from each individual eigenvalue and then dividing by the standard deviation or other metric.
[0094] As another example of a preprocessing technique, some or all of the input data may be quantized or discretized. In some cases, qualitative features or variables contained in the input data may be converted to quantitative features or variables.
[0095] Figure 3A 312 is a block diagram illustrating an example ear-worn device modeled by an example machine learning model according to one or more aspects of the present disclosure. Ear-worn device 312 may be any type of ear-worn device configured to provide auditory sounds. Ear-worn device 312 is a device designed to be worn on or in a user's ear.
[0096] The ear-worn device 312 includes one or more communication units 314, one or more input components 318, one or more output components 320, one or more processors 322, and a battery / charging circuit 326 operatively or communicatively coupled via a communication bus 324. Each of the aforementioned components of the ear-worn device 312 is shown as being housed within a housing 330, including within the housing 330, a vent 332, and an access panel 334.
[0097] Similar to Figure 2 The communication unit 314 may include a radio or other component that enables the ear-worn device 312 to communicate data with other devices. For example, the communication unit 314 may enable the ear-worn device to exchange data with another ear-worn device (e.g., for a user with bilateral hearing loss, for a set of stereo headphones, etc.). Similarly, the communication unit 314 may enable the ear-worn device to exchange data with a programmer, mobile phone, or other external device.
[0098] The battery / charging circuit 326 of the ear-worn device 312 provides power to various components of the ear-worn device 312 to enable the ear-worn device 312 to provide sound. Figure 2 The processor 322 may perform operations that enable the ear-worn device 312 to perform functions or take specific actions. As an example, the processor 322 may include a digital signal processor or other processing circuit that controls various components of the ear-worn device 312 to provide auditory sounds.
[0099] The ear-worn device 312 includes one or more input components 318 and one or more output components 320. The input component 318 may include a microphone configured to detect audible sounds in a user's listening area, and the output component 320 may include a receiver (e.g., a speaker) configured to generate auditory sounds, for example, based on the detected audible sounds.
[0100] The vent 332 of the ear-worn device 312 provides a conduit through which sound can enter and exit the housing 330. For example, the microphone of the input component 318 can be configured to receive sound waves via the vent 332, and the receiver of the output component 320 can be configured to output sound waves via the vent 332. The vent 332 can have a specific shape and / or size that is specifically designed to accommodate the housing 330 for inputting and outputting sound.
[0101] The access panel 334 of the ear-worn device 312 provides a way for the user to access the interior cavity 332 of the ear-worn device 312, for example, to replace the battery / charging circuit 326, or to access other components of the ear-worn device 312. The access panel 334 can be a door, window, or other removable cover.
[0102] A set of empirical modeling guidelines known to the ear-worn device manufacturer can specify a preferred layout for each component of the ear-worn device 332. For example, the guidelines can define a minimum thickness of the housing 330. The guidelines can further specify the location of the access panel 334 relative to the other components of the ear-worn device 332. Similarly, the set of empirical modeling guidelines can define minimum separation distances from various components of the ear-worn device 312 and the housing 330, for example, by specifying that a receiver of the output component 320 should not touch the housing 330 or be within a minimum distance threshold of the housing 330. Other design considerations and guidelines may follow. Reference Figure 1 , these empirical modeling guidelines can be quantified and used as training data 104 or input data 106 .
[0103] Figure 3B is a conceptual diagram illustrating a cross-sectional view of a representation of a target ear canal including a position along an axis protruding through the target ear canal according to one or more aspects of the present disclosure. Figure 2 The context of computing device 212 is described in Figure 3B For example, in accordance with the techniques of this disclosure, ML model 202 of computing device 212 may receive as input a representation 336 of a series of three-dimensional points defining an ear canal of a user, illustrated as target ear canal 340 , and generate a representation of a wearable ear device for target ear canal 340 .
[0104] The representation 336 of the target ear canal 340 includes an ear opening at one end and an ear drum at an opposite end. The representation 336 of the target ear canal 340 is defined in part by an axis 342 protruding from the ear drum at one end of the target ear canal 340 to the opposite end of the target ear canal 340.
[0105] ML model 202 can define axis 342 to include multiple positions 344A-344E protruding from an ear drum of target ear canal 340 to an ear opening at an opposite end of target ear canal 340. Position 344E is shown closest to the ear opening and farthest from the ear drum, and position 344A is shown closest to the ear drum and farthest from the ear opening. ML model 202 can generate a representation of the ear-worn device to minimize an error in the representation of the ear-worn device, wherein the error is a function of the position and curvature along axis 342.
[0106] For example, training module 226 of computing device 212 can determine deviations between expected outputs of ML model 202 and actual outputs of ML model 202. Training module 226 can cause ML model 202 to place greater emphasis on correcting certain deviations than other models. For example, training module 226 can cause ML model 202 to apply higher weights to ensure that the representation of the ear-worn device is most accurate in certain regions along axis 342, and in some examples, ML model 202 can apply lower weights to other regions along axis 342 that may be less important to ensure usability and user comfort.
[0107] For example, locations 344D and 344B are located in areas of ear canal 340 having greater curvature above and below axis 342 than locations 344A, 344C, and 344E, where ear canal 340 has less curvature above and below axis 342. User discomfort may occur near locations 344D and 344B, for example, more than at locations 344A, 344C, and 344E if the housing of the earworn is not properly shaped and molded to fit within the curvature of these areas.
[0108] The training module 226 can apply respective weights to the locations 344A-344E along the axis 342. The respective weights for each of the different locations 344A-344E can be a function of the amount of curvature at the location. The ML model 202 can shape the portion of the housing of the ear-worn device at the location with a larger respective weight along the axis 342 to be more accurate to the target ear canal 340 than other locations along the axis 342 with a lower respective weight.
[0109] For example, the training module 226 can determine the deviation of the shape of the ear-worn device compared to the desired or expected shape. The training module 226 can apply weights to more important areas in the ear-worn device housing where the deviation occurs. The training module 226 can apply weights to cause the ML model 202 to emphasize reducing the shell shape error at locations 344D and 344B rather than reducing the shell shape error at other locations 344A, 344C, and 344E. The training module 226 can increase the corresponding weights of locations 344B and 344D having a larger amount of curvature, and can reduce the corresponding weights of locations 344A, 344C, and 344E having a smaller amount of curvature.
[0110] Figure 4 is a flow chart illustrating operations performed by an example machine learning system according to one or more aspects of the present disclosure. Figure 2 The computing device 212 is described in the context of Figure 4 The computing device 212 may be configured in a manner different from Figure 4 The operations 402 to 416 are performed in the order shown. The computing device 212 may perform the comparison Figure 4 More or fewer operations may be performed as shown.
[0111] In operation, computing device 212 may receive representations of previously manufactured ear-worn devices from a production facility database and / or a clinician / retail store database (402). For example, training module 226 of computing device 212 may communicate with a remote computing device using communication unit 214 that maintains a database or other record of previously manufactured ear-worn devices. Training module 226 may receive data from a production facility database or a clinician or retail store database in its finished form, the data including representations of previously manufactured ear-worn devices, and based on the production data, create training data to train ML model 202 to replicate ear-worn devices having features similar to the previously manufactured ear-worn devices. Additionally or alternatively, training module 226 may receive other data including representations of previously manufactured ear-worn devices from a clinician or retail store database, and based on the clinician or retail data, create training data to train ML model 202 to replicate ear-worn devices having features similar to the previously manufactured ear-worn devices.
[0112] The computing device 212 may train the ML model based on a representation of a previously manufactured ear-worn device and a representation of a corresponding ear impression used to generate the previously manufactured ear-worn device (404). For example, for a target ear canal impression, the desired output or training target of the ML model 202 may be a set of three-dimensional points or a Fourier transform thereof that defines a final ear-worn device housing. The cost function used by the training module 226 to train the ML model 202 may indicate the deviation between the output of the ML model 202 and the desired output (e.g., a weighted mean square error). When the ear-worn device is located in the target ear canal, the actual weight applied to the deviation associated with different parts of the ear-worn device housing may be a function of the housing curvature and relative position (e.g., to focus the ML model 202 on generating an ear-worn device housing design that is comfortable to wear). The training module 226 may utilize a back-propagation algorithm to further train the ML model 202.
[0113] By accessing records of previously manufactured ear-worn devices produced at production facilities and / or clinics or retail stores, training module 226 can train ML model 202 to mimic ear-worn shell designs at any stage of the product life cycle, including during production and customer delivery. For example, in addition to having ML model 202 learn what a skilled modeler does in production, training module 226 can input training data derived from a clinic so that ML model 202 learns and mimics not only what a modeler does in production, but also what a professional does in the clinic. In other words, instead of using the final shell design in production as a training target, training module 226 can use the final shell design modified by a clinician as a training target.
[0114] In some examples, computing device 212 can further train the ML model based on characteristics of the user or other information associated with previously manufactured ear-worn devices (406). For example, training module 226 can adapt the training of model 202 to various other conditions that may affect ear-worn housing design, shape, and preferences. For example, training module 226 can use a specific training data set (e.g., in training data storage 204) associated with a specific country, a specific region, or even a specific clinic, instead of training ML model 202 to be generally configured for the entire user population. In other words, users in a specific country, region, clinic, or other demographic may have specific preferences; training module 226 can train ML model 202 to generate representations of ear-worn devices that adapt to these preferences.
[0115] The computing device 212 can obtain a representation of the target ear canal ( 408 ).For example, the UI module 228 can receive a scan of the end user's ear canal and provide the scan (or its Fourier transform) as input to the ML model 202 .
[0116] The computing device 212 can input the representation of the target ear canal and any other relevant inputs to the ML model to generate a representation of the ear-worn device for the target ear canal (410). For example, the UI module 228 can receive other user inputs specifying a geographic region, body type, skin type, or other characteristics associated with the user or the ordering clinic, the other user inputs specifying preferences for ear-worn device designs. The UI module 228 can format and send the various inputs to the model 202, including the representation of the target ear canal.
[0117] Computing device 212 may output an indication of a representation of the ear-worn device for the target ear canal (412). For example, ML model 202 may output data indicating a shape and configuration of the ear-worn device. In some examples, the output of ML model 202 may define a set of points or a Fourier transform thereof that define a housing of the ear-worn device. The output of ML model 202 may specify locations of various internal components within the housing.
[0118] The computing device 212 can receive user feedback associated with the ear-worn device for the target ear canal (414). For example, a skilled modeler can make improvements or modifications to the final assembled ear-worn device based on the output from the model 202. Any adjustments or modifications made by the modeler can be determined by the UI module 228 and sent to the training module 226.
[0119] Computing device 212 can train the ML model based on user feedback associated with the ear-worn device for the target ear canal (416). For example, training module 226 can input representations of modifications made to outputs from ML model 202 so that ML model 202 learns and anticipates these modifications when generating future ear-worn device designs. In some cases, ML model 202 can be continuously trained and improved as additional training data and shell modeling guidelines are developed, for example, as new ear-worn device technologies and designs continue to develop.
[0120] It should be appreciated that, depending on the examples, certain actions or events of any of the techniques described herein may be performed in a different order, may be added, combined, or omitted altogether (e.g., not all described actions or events are necessary for the practice of the techniques). In addition, in some examples, actions or events may be performed simultaneously, e.g., via multithreading, interrupt handling, or multiple processors, rather than sequentially.
[0121] In one or more examples, the functions described may be implemented in hardware, software, firmware, or any combination thereof. If implemented in software, the functions may be stored on or transmitted via a computer-readable medium as one or more instructions or codes and executed by a hardware-based processing unit. A computer-readable medium may include a computer-readable storage medium corresponding to a tangible medium such as a data storage medium or a communication medium including, for example, any medium that facilitates the transfer of a computer program from one place to another according to a communication protocol. In this manner, a computer-readable medium may generally correspond to (1) a non-temporary tangible computer-readable storage medium, or (2) a communication medium such as a signal or carrier wave. A data storage medium may be any available medium that can be accessed by one or more computers or one or more processing circuits to retrieve instructions, codes, and / or data structures to implement the techniques described in the present disclosure. A computer program product may include a computer-readable medium.
[0122] By way of example and not limitation, such computer-readable storage media may include RAM, ROM, EEPROM, CD-ROM or other optical disk storage, disk storage or other magnetic storage device, flash memory, cache memory or any other medium that can be used to store desired program code in the form of instructions or data structures and can be accessed by a computer. In addition, any connection can be considered as a computer-readable medium. For example, if a coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL) or wireless technology (e.g., infrared, radio and microwave) is used to transmit instructions from a website, server or other remote source, coaxial cable, fiber optic cable, twisted pair, DSL or wireless technology (e.g., infrared, radio and microwave) are included in the definition of medium. However, it should be understood that computer-readable storage media and data storage media do not include connection, carrier wave, signal or other temporary media, but point to non-temporary tangible storage media. The above combination should also be included in the scope of computer-readable media.
[0123] The functions described in this disclosure may be performed by fixed-function and / or programmable processing circuits. For example, instructions may be performed by fixed-function and / or programmable processing circuits. Such processing circuits may include one or more processors, for example, one or more digital signal processors (DSPs), general-purpose microprocessors, application-specific integrated circuits (ASICs), field programmable logic arrays (FPGAs), or other equivalent integrated or discrete logic circuits. Therefore, the term "processor" used herein may refer to any of the aforementioned structures or any other structure suitable for implementing the technology described herein. In addition, in some aspects, the functions described herein may be provided in dedicated hardware and / or software modules. Similarly, these technologies may be implemented entirely in one or more circuits or logic elements. The processing circuit may be coupled to other components in various ways. For example, the processing circuit may be coupled to other components via internal device interconnection, wired or wireless network connection, or another communication medium.
[0124] Various components, modules, or units are described in this disclosure to emphasize functional aspects of devices configured to perform the disclosed techniques, but do not necessarily need to be implemented by different hardware units. Instead, as described above, the various units can be combined in hardware units or provided by a collection of interoperable hardware units including one or more processors as described above and appropriate software and / or firmware.
[0125] Various examples have been described. These and other examples are within the scope of the following claims.
[0126] The present invention can also be configured as follows.
[0127] 1. A method comprising:
[0128] obtaining, by at least one processor, a representation of a target ear canal of a user; and
[0129] A machine learning model is applied by the at least one processor to generate a representation of the ear-worn device of the target ear canal, wherein inputs to the machine learning model include the representation of the target ear canal of the user and representations of one or more internal components of the ear-worn device.
[0130] 2. The method of claim 1, wherein the one or more internal components of the ear-worn device include one or more of: a receiver, a microphone, or a circuit.
[0131] 3. The method of claim 1 , wherein the one or more internal components of the ear-worn device include one or more of a battery or a charging circuit.
[0132] 4. The method of claim 1, wherein the one or more internal components of the ear-worn device include a vent.
[0133] 5. The method of claim 1 , wherein the representation of the one or more internal components of the ear-worn device comprises physical dimensions of the one or more internal components of the ear-worn device.
[0134] 6. The method of claim 1, wherein the input to the machine learning model further comprises one or more constraints on the position of one or more internal components of the ear-worn device.
[0135] 7. The method of claim 6, wherein the one or more limitations include a minimum distance between an internal component of the one or more internal components and a housing of the ear-worn device.
[0136] 8. The method of 1, wherein the machine learning model is a deep learning model.
[0137] 9. The method of claim 1, wherein the machine learning model is trained based at least in part on representations of previously manufactured ear-worn devices and representations of ear-worn devices in an ear canal.
[0138] 10. The method according to 1 further includes: training the machine learning model based on feedback from an operator, wherein the feedback includes changes in the size of the shell of the ear-worn device.
[0139] 11. The method according to 1 further includes: training the machine learning model based on feedback from an operator, wherein the feedback includes changes in the position of one or more internal components located within the shell of the ear-worn device.
[0140] 12. The method of 1, wherein, after training the machine learning model, generating the representation of the ear-worn device comprises: further inputting one or more of the following into the machine learning model:
[0141] an indication of a specific type of hearable; or
[0142] An indication of one or more modeling constraints.
[0143] 13. The method of claim 1, wherein the representation of the target ear canal comprises a Fourier transform of the representation of the target ear canal.
[0144] 14. The method of claim 1, wherein, for each of the one or more internal components, the representation of the internal component comprises a three-dimensional representation of the internal component or a Fourier transform of the three-dimensional representation of the internal component.
[0145] 15. A non-transitory computer-readable storage medium comprising instructions that, when executed by at least one processor, cause the at least one processor to:
[0146] obtaining a representation of a target ear canal of a user; and
[0147] Applying a machine learning model to generate a representation of the ear-worn device for the target ear canal, wherein inputs to the machine learning model include the representation of the target ear canal of the user and representations of one or more internal components of the ear-worn device.
[0148] 16. The non-transitory computer-readable storage medium of 15, wherein the representation of the one or more internal components of the ear-worn device comprises physical dimensions of the one or more internal components of the ear-worn device.
[0149] 17. The non-transitory computer-readable storage medium of 15, wherein the input to the machine learning model further comprises one or more constraints on the position of the one or more internal components of the ear-worn device.
[0150] 18. The non-transitory computer-readable storage medium of 17, wherein the one or more limitations include a minimum distance of an internal component of the one or more internal components from a housing of the ear-worn device.
[0151] 19. A computing device comprising:
[0152] One or more data storage devices configured to store the machine learning model; and
[0153] One or more processors configured to:
[0154] obtaining a representation of a target ear canal of a user; and
[0155] Applying a machine learning model to generate a representation of the ear-worn device for the target ear canal, wherein inputs to the machine learning model include the representation of the target ear canal of the user and representations of one or more internal components of the ear-worn device.
[0156] 20. The computing device of 19, wherein the one or more internal components of the ear-worn device include one or more of: a receiver, a microphone, or a circuit.
Claims
1. A method for modeling a housing of an ear-worn device, comprising: Training a machine learning model to generate a representation of an ear-worn device for an ear canal, wherein training the machine learning model comprises: receiving a data set including a representation of a historical ear impression and a corresponding target representation of an ear-worn device; providing the historical ear impression to the machine learning model, wherein the machine learning model is configured to output a generated representation of the ear-worn device; applying a cost function to an output of the machine learning model to determine a cost, wherein the cost function comprises a weighted mean squared error of a deviation of the generated representation of the ear-worn device and the target representation of the ear-worn device; and updating a model parameter of the machine learning model based on the cost; receiving, by at least one processor, a representation of a target ear canal of a patient; and The machine learning model is applied by the at least one processor to generate a representation of the ear-worn device of the target ear canal, wherein inputs to the machine learning model include the representation of the target ear canal of the patient and representations of one or more internal components of the ear-worn device.
2. The method according to claim 1, wherein: The one or more internal components of the ear-worn device include one or more of the following: a receiver, a microphone, or a circuit.
3. The method according to claim 1, wherein: The one or more internal components of the ear-worn device include one or more of a battery or a charging circuit.
4. The method according to claim 1, wherein: The one or more internal components of the ear-worn device include a vent.
5. The method according to claim 1, wherein: The representation of the one or more internal components of the ear-worn device includes physical dimensions of the one or more internal components of the ear-worn device.
6. The method according to claim 1, wherein: The input to the machine learning model further includes one or more constraints on the position of the one or more internal components of the ear-worn device.
7. The method according to claim 6, wherein: The one or more constraints include a minimum distance of an internal component of the one or more internal components from a housing of the ear-worn device.
8. The method according to claim 1, wherein: The machine learning model is a deep learning model.
9. The method according to claim 1, further comprising: The machine learning model is trained based on feedback from an operator, the feedback including dimensional changes of a shell of the ear-worn device.
10. The method according to claim 1, further comprising: The machine learning model is trained based on feedback from an operator, the feedback including changes in positions of the one or more internal components located within a shell of the ear-worn device.
11. The method according to claim 1, wherein: After training the machine learning model, applying the machine learning model to generate the representation of the ear-worn device includes: further inputting one or more of the following into the machine learning model: an indication of a specific type of hearable; or An indication of one or more modeling constraints.
12. The method according to claim 1, wherein: The representation of the target ear canal comprises a Fourier transform of the representation of the target ear canal.
13. The method according to claim 1, wherein: For each internal component of the one or more internal components, the representation of the internal component comprises a three-dimensional representation of the internal component or a Fourier transform of the three-dimensional representation of the internal component.
14. A non-transitory computer-readable storage medium comprising instructions that, when executed by at least one processor, cause the at least one processor to: A machine learning model is trained to generate a representation of an ear-worn device for an ear canal, wherein: The at least one processor is configured, as part of training the machine learning model: receiving a data set including a representation of a historical ear impression and a corresponding target representation of an ear-worn device; providing the historical ear impression to the machine learning model, wherein the machine learning model is configured to output a generated representation of the ear-worn device; applying a cost function to an output of the machine learning model to determine a cost, wherein the cost function comprises a weighted mean squared error of a deviation of the generated representation of the ear-worn device and the target representation of the ear-worn device; and updating a model parameter of the machine learning model based on the cost; receiving a representation of a target ear canal of the patient; and Applying the machine learning model to generate a representation of the ear-worn device of the target ear canal, wherein inputs to the machine learning model include the representation of the target ear canal of the patient and representations of one or more internal components of the ear-worn device.
15. The non-transitory computer-readable storage medium of claim 14, wherein: The representation of the one or more internal components of the ear-worn device includes physical dimensions of the one or more internal components of the ear-worn device.
16. The non-transitory computer-readable storage medium of claim 14, wherein: The input to the machine learning model further includes one or more constraints on the position of the one or more internal components of the ear-worn device.
17. The non-transitory computer-readable storage medium of claim 16, wherein: The one or more constraints include a minimum distance of an internal component of the one or more internal components from a housing of the ear-worn device.
18. A computing device comprising: one or more data storage devices configured to store the machine learning model; as well as One or more processors configured to: training the machine learning model to generate a representation of the ear-worn device for the ear canal, wherein the one or more processors are configured, as part of training the machine learning model: receiving a data set including a representation of a historical ear impression and a corresponding target representation of an ear-worn device; providing the historical ear impression to the machine learning model, wherein the machine learning model is configured to output a generated representation of the ear-worn device; applying a cost function to an output of the machine learning model to determine a cost, wherein the cost function comprises a weighted mean squared error of a deviation of the generated representation of the ear-worn device and the target representation of the ear-worn device; and updating a model parameter of the machine learning model based on the cost; receiving a representation of a target ear canal of a user; and Applying the machine learning model to generate a representation of the ear-worn device of the target ear canal, wherein inputs to the machine learning model include the representation of the target ear canal of the user and representations of one or more internal components of the ear-worn device.
19. The computing device of claim 18, wherein: The one or more internal components of the ear-worn device include one or more of the following: a receiver, a microphone, or a circuit.
Citation Information
Patent Citations
Backup steering system for steer-by-wire harvesters
US20190061807A1
Self-fitting in-canal component for hearing aids
CN106941652A
Custom earbud scanning and fabrication
CN108476352A
Method for modelling customised earpieces
US20040107080A1
System and method for performing a selective fill for a hearing aid shell
US20080094392A1
