Intelligent hearing threshold evaluation method and system

Abstract

The application discloses an intelligent hearing threshold evaluation method and system. The method comprises the following steps: in the initialization stage, compensation parameters are pushed and real-time compensation is carried out according to the acoustic characteristics of the terminal device; environmental noise is collected in real time before the test sound is played, and the noise masking threshold and the dynamic gain compensation amount are calculated; the reference hearing level is obtained according to the test protocol, the target sound pressure level is obtained in combination with the compensation parameters and the dynamic gain compensation amount, the pure tone test sound is played after being converted into a gain value; the sound is collected and the time stamp is recorded; the sound signal is subjected to streaming speech recognition, signal separation is carried out through time-frequency mask technology and a confidence score is output, and the hearing level is determined in combination with the effective response. The application can dynamically analyze the noise characteristics and adaptively adjust the sound intensity compensation, eliminate the device difference, and improve the accuracy and accessibility of hearing evaluation in non-professional scenes.

Description

Technical Field

[0001] This invention relates to an intelligent hearing threshold assessment method and a corresponding intelligent hearing threshold assessment system, belonging to the field of hearing assessment technology. Background Technology

[0002] Hearing impairment is a significant health issue affecting the quality of life for older adults. Those with hearing impairment not only experience difficulties in verbal communication but may also experience memory decline, an increased risk of Alzheimer's disease, and psychological problems such as anxiety and loneliness. Early hearing screening is crucial for timely detection of hearing impairment and for halting disease progression and enabling early intervention. Therefore, developing convenient and accurate hearing assessment methods is of great practical significance for primary healthcare, community settings, and home environments.

[0003] Existing hearing screening technologies for the elderly are mainly divided into two categories. One category is hearing tests, including clinical methods such as pure-tone audiometry and speech audiometry; the other category is hearing scale assessments, such as the "Hearing Screening Scale for the Elderly" and other subjective assessment tools. With the development of mobile internet technology, the World Health Organization launched the hearing assessment software "hear WHO". In this software, subjects access the application through mobile devices such as smartphones. The application generates standard test audio (such as digital speech), which is played through various types of headphones worn by the subjects. Subjects input feedback on the interface based on the audio content they hear (such as digit sequences). The program adjusts the volume of the digital audio based on the responses, and after answering a certain number of questions, it displays the subject's hearing level. This type of mobile hearing assessment tool breaks through the dependence of traditional testing on professional equipment and venues, and has good potential for widespread adoption.

[0004] However, existing mobile hearing assessment technologies still face many challenges in practical applications. First, the tests require a strictly quiet environment; ambient noise in non-professional settings such as homes and communities can easily mask the test signal, leading to inaccurate threshold determinations. Second, different brands and models of consumer-grade terminal devices (such as smartphones and headphones) exhibit significant differences in maximum output sound pressure level, frequency response curves, and total harmonic distortion. This can result in the same digital audio file being played on different devices, leading to a difference of ±15 dB or more in the actual sound pressure level received in the user's ear canal, making the test results incomparable. Furthermore, the user interfaces of existing assessment software are mostly generic designs, failing to fully consider the specific needs of elderly individuals in hearing threshold assessment operations. Issues such as small fonts, unclear buttons, and misinterpretations of unclear pronunciation by speech recognition cause some elderly individuals to be unable to complete the entire assessment task independently, or even give up halfway through.

[0005] The combined effect of these issues means that current mobile hearing assessment processes often require one-on-one guidance from professionals. This significantly increases the labor costs for primary healthcare institutions and limits the screening coverage in non-professional settings such as communities and homes, substantially reducing the accessibility and accuracy of hearing health screenings for the elderly. Therefore, how to achieve accurate, convenient, and autonomous hearing threshold assessment in complex environments, with diverse equipment, and for elderly users is a pressing technical problem that needs to be solved. Summary of the Invention

[0006] The primary technical problem to be solved by this invention is to provide an intelligent hearing threshold assessment method.

[0007] Another technical problem to be solved by the present invention is to provide an intelligent hearing threshold assessment system.

[0008] To achieve the above-mentioned technical objectives, the present invention adopts the following technical solution: According to a first aspect of the present invention, an intelligent hearing threshold assessment method is provided, comprising the following steps: S1: During the initialization phase, compensation parameters are pushed according to the acoustic characteristics of the terminal device, and the compensation parameters are used to compensate the playback sound of the terminal device in real time to ensure that different terminal devices output the same physical sound pressure level. S2: During the testing phase, noise is collected in real time before the test sound is played, which is used to calculate the noise masking threshold and dynamic gain compensation amount. S3: Obtain the reference hearing level corresponding to the current test frequency according to the test protocol; use the compensation parameters obtained in step S1 to convert the reference hearing level into a reference sound pressure level, and use the dynamic gain compensation amount obtained in step S2 to compensate the reference sound pressure level to obtain the compensated target sound pressure level. S4: The gain value is calculated based on the target sound pressure level and acoustic characteristics and is used to test sound playback; S5: Play a pure tone test tone and record the playback start timestamp at the same time; S6: Collect audio during the test and record timestamps; S7: Perform streaming speech recognition on the sound signal to obtain an effective response, and simultaneously perform signal separation to output a confidence score. Combine the effective response and the confidence score to determine the subject's hearing level.

[0009] Preferably, in step S1, the compensation parameters are obtained by querying a mapping table and matching the compensation parameters {a, b, c, frequency response curve}, where a represents the sensitivity coefficient of the sound played by the terminal device; b represents the frequency response slope compensation coefficient; and c represents the reference offset constant. The query mapping table contains the device model and the frequency response characteristic parameters that match each device model.

[0010] Preferably, in step S7, the Transformer time-frequency masking algorithm is used to separate noise from the potential speech / test signal to ensure that noise feature extraction is not contaminated by sudden interference.

[0011] A preferred approach is to use a multi-channel IMFCC feature extraction algorithm to process multiple feature streams with different time-frequency resolutions or different frequency bands of interest in parallel, perform millisecond-level spectrum tracking of noise intensity and frequency characteristics, and construct a three-dimensional noise feature model including intensity, frequency, and time modulation characteristics to effectively distinguish between steady-state and transient noise.

[0012] Preferably, the multi-channel refers to multiple feature channels constructed using a multi-resolution parallel analysis method to address the different time scale, frequency scale, and modulation scale characteristics of environmental noise.

[0013] Preferably, the intelligent hearing threshold assessment method further includes step S8: collecting data and environmental noise from all test sessions, updating the three-dimensional noise feature model, and continuously updating the noise database and the acoustic characteristic database of the terminal device.

[0014] Preferably, dynamic spectrum enhancement technology is used to adaptively adjust the gain compensation amount of the test audio based on the real-time noise fingerprint, so that the actual perceived signal intensity is at least higher than the target signal-to-noise ratio than the noise masking threshold; wherein, the gain compensation amount is the smaller value between the minimum gain required to overcome the masking and the maximum safe gain allowed by the device, and is not less than 0.

[0015] Preferably, in step S3, the reference hearing level specified in the test protocol is obtained by nonlinear mapping to obtain the reference sound pressure level; then, the gain compensation amount calculated in real time in step S2 is used for compensation to obtain the compensated target sound pressure level.

[0016] A preferred approach is to use a hyperbolic tangent mapping function to adapt to auditory nonlinearity, thereby converting the reference hearing level to the reference sound pressure level.

[0017] According to a second aspect of the present invention, an intelligent hearing threshold assessment system is provided, comprising: Terminal devices are used to collect environmental noise, play test audio, pick up user responses, and display evaluation results; A cloud server is used to implement the steps in the above-mentioned intelligent hearing threshold assessment method. The terminal device and the cloud server interact with each other via a wireless communication network. The cloud server sends compensation parameters based on a lookup mapping table of the device model reported by the terminal device; calculates dynamic gain compensation based on the environmental noise collected in real time by the terminal device; executes streaming speech recognition and Transformer time-frequency masking algorithms, outputs a confidence score, and determines the hearing level. The terminal device then performs sound pressure level compensation on the test audio based on the compensation parameters and dynamic gain compensation sent by the cloud server before playing it.

[0018] Compared with existing technologies, the intelligent hearing threshold assessment method and system provided by this invention eliminates hardware differences in frequency response curves, sensitivity, etc., between different consumer-grade devices by calibrating and compensating the acoustic characteristics of terminal devices in real time. This ensures that the same test audio can output a consistent physical sound pressure level on different devices, thereby significantly improving the comparability and accuracy of cross-device test results. Simultaneously, this invention collects ambient noise in real time before each test audio playback, accurately constructs a three-dimensional noise feature model using multi-channel feature extraction and time-frequency masking technology, and adaptively calculates gain compensation based on a dynamic spectrum enhancement algorithm. This effectively overcomes environmental noise interference in non-professional scenarios such as homes and communities, ensuring that the test signal can be stably transmitted to the user's ear canal with a signal-to-noise ratio better than the masking threshold. Furthermore, this invention combines streaming speech recognition, confidence scoring, and intelligent interactive guidance, reducing the difficulty of operation and error rate for the elderly, enabling self-testing without one-on-one guidance from professionals. In summary, this invention significantly improves the stability, consistency, accessibility, and user compliance of hearing threshold assessment in non-clinical scenarios. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall process of the intelligent hearing threshold assessment method provided in the embodiments of the present invention; Figure 2 for Figure 1 The flowchart of the steps shown in the double-dotted box (i.e., the steps involved in the testing phase, such as noise acquisition, dynamic gain compensation, test sound playback, speech recognition and confidence determination) is shown in the middle. Figure 3 This is a schematic diagram of the appearance of the terminal device (headphones) used in the embodiments of the present invention. Detailed Implementation

[0020] The technical content of the present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0021] First Embodiment like Figures 1 to 3 As shown, the first embodiment of the present invention provides an intelligent hearing threshold assessment method, which, after receiving an assessment request sent by a user through a terminal device (i.e., a sub-terminal), performs the following steps: S1: During the initialization phase, compensation parameters are pushed based on the acoustic characteristics of the terminal device, and these compensation parameters are used to compensate the playback sound of the terminal device in real time to ensure that different devices output the same physical sound pressure level.

[0022] When the subject (user) logs into the cloud server using a terminal device, the terminal device enters the initialization phase.

[0023] It should be noted that the terminal devices used in the embodiments of the present invention have two implementation methods: one is to use dedicated equipment, and the other is to use conventional equipment (i.e., mobile phones of various brands and models on the market). Here, the method of using dedicated equipment will be described first.

[0024] In one embodiment of the present invention, the dedicated device is a headset with a built-in active noise cancellation unit, which can specifically shield environmental interference to improve the accuracy of the evaluation. This headset, used in conjunction with a conventional tablet or smartphone, enables the technical solution of the present invention. The headset serves as both a sound output device and a sound pickup device; the tablet or smartphone serves as a local terminal device, possessing display functionality, data processing capabilities, and the ability to communicate with a cloud server.

[0025] The cloud server pre-stores acoustic characteristic databases and noise databases for the terminal devices. When constructing the acoustic characteristic database for the terminal devices, in a laboratory environment, professional equipment such as artificial ears and acoustic testing chambers are used to measure and evaluate parameters such as frequency response curves, sensitivity, and total harmonic distortion of different models of PADs and dedicated headphones. A lookup mapping table between device models and calibration parameters is established and entered into the acoustic characteristic database of the terminal devices. During evaluation, a standard swept-frequency signal (125 Hz to 8 kHz, step 1 / 3 octave) is played through the device's built-in speaker (or headphone unit), and the playback signal is picked up by a microphone. This signal is used to calculate the difference between the actual output sound pressure level and the target sound pressure level, generating personalized compensation parameters for the device.

[0026] The pre-entered query mapping table contains the device model and its matching frequency response parameters and compensation parameters. For example, different headphones often have different gains at 1 kHz and 4 kHz, so it is necessary to generate corresponding compensation parameters based on the frequency response parameters (i.e., frequency response curves) of the terminal device. The WebAudio application on the local terminal device compensates for the played sound based on these compensation parameters, thereby ensuring that different devices output the same physical sound pressure level.

[0027] Specifically, based on the acoustic characteristic database of the terminal device, during the initialization phase of each hearing threshold test, the terminal device reports its device model to the cloud server. The cloud server matches the compensation parameters {a, b, c, frequency response curve} according to the device model lookup mapping table and sends them to the terminal device. The frequency response curve is the actual measured output physical sound pressure level (LdB) of the terminal device at different frequencies f. The three coefficients used to describe the linear response portion of this frequency response curve are: a - the sensitivity coefficient of the sound played by the terminal device, i.e., the ratio of the output sound pressure level to the input digital signal; b - the frequency response slope compensation coefficient; and c - the reference offset constant. These three parameters are collectively referred to as the "device physical response calibration coefficients," used to describe the linear acoustic response characteristics of the terminal device (headphones or tablet) in order to convert the physical sound pressure level (LdB) into physical playback amplitude (i.e., DAC output value).

[0028] The noise database contains common noise types: white noise, pink noise, air conditioner noise, human voices, vehicle noises, door opening and closing sounds, etc. A pre-trained Transformer time-frequency masking neural network is used to calculate the noise masking threshold during the initialization phase and dynamically updates the noise database in closed-loop self-iteration.

[0029] S2: During the testing phase, before the test sound is played, the ambient noise sensor of the terminal device is used to collect noise in real time to calculate the noise masking threshold and dynamic gain compensation (ΔdB).

[0030] After receiving the evaluation task pushed by the cloud server, the environmental noise sensor on the terminal device (e.g., headphones) continuously collects noise data of the test scene (sampling rate ≥ 44.1 kHz, bit depth ≥ 16 bit) and uploads it to the local terminal device. The terminal device then transmits the noise data stream to the cloud server via wireless communication methods such as 3G, 5G / Wi-Fi.

[0031] The cloud server invokes a multi-channel IMFCC (Inverse Mel-Frequency Cepstral Coefficients) algorithm, configures parameters such as filter banks, window length, and frame shift, and initiates data stream acquisition from the environmental noise sensor for millisecond-level spectrum tracking. This extracts the time-frequency domain features of the noise, constructing a three-dimensional noise feature model that includes noise intensity (Leq), frequency distribution (1 / 3 octave spectrum), and time modulation characteristics (steady-state / transient). Simultaneously, a Transformer time-frequency masking algorithm is used to separate noise from the potential speech / test signal, ensuring that noise feature extraction is not contaminated by sudden interference (such as door slamming or coughing). Based on the three-dimensional noise feature model, steady-state noise (such as air conditioning or fan noise) and transient noise (such as conversation or traffic noise) are dynamically distinguished, and their energy proportions are labeled accordingly. Therefore, by effectively distinguishing between steady-state and transient noise, dynamic spectrum enhancement technology allows the cloud server to dynamically analyze noise intensity and frequency characteristics, automatically adjusting the sound intensity compensation parameters of the test audio (adaptively adjusting gain based on real-time noise fingerprints), ensuring that the test signal can still be effectively transmitted in complex environments.

[0032] The cloud server employs a combination of dynamic algorithms and real-time processing, utilizing a multi-channel IMFCC (Inverse Mel Frequency Cepstral Coefficient) feature extraction algorithm to process multiple feature streams with different time-frequency resolutions or frequency bands of interest in parallel. This enables millisecond-level spectral tracking of noise intensity and frequency characteristics, accurately constructing a three-dimensional noise feature model (intensity + frequency + time modulation characteristics) to effectively distinguish between steady-state and transient noise. Unlike ordinary FFT, which can only capture intensity and frequency features, IMFCC can also capture time modulation characteristics (such as the fluctuation pattern of noise), which can be used to distinguish between steady-state noise (e.g., air conditioner noise) and transient noise (e.g., door closing sound). During multi-channel parallel processing, each channel outputs its own feature tensor, which is then merged into the final three-dimensional noise feature model through a feature fusion layer (e.g., weighted averaging, attention mechanisms, or splicing). While conventional IMFCC has only one set of parameters, the multi-channel IMFCC in this embodiment captures multi-scale features through parallel configurations to construct the initial three-dimensional noise feature model. For example, Channel 1 (low-frequency high resolution) processes audio signals from 0 to 4 kHz with a window length of 512 points; Channel 2 (mid-frequency balanced) processes audio signals from 2 kHz to 12 kHz with a window length of 256 points; and Channel 3 (high-frequency high-time precision) processes audio signals from 4 kHz to 24 kHz with a window length of 128 points. The output three-dimensional noise feature model includes: 1) Intensity features: equivalent sound pressure level, band energy, percentile; 2) Frequency features: inverse Mel-frequency cepstral coefficients, spectral centroid, spectral roll-off point; and 3) Time modulation features: modulation spectrum, steady-state / transient score, and event density. These features are extracted and used as a noise fingerprint.

[0033] It should be noted that the multi-channel processing in this embodiment of the invention is not a simple multiplexed process, but rather a multi-resolution parallel analysis of multiple feature channels designed to address the different time-scale, frequency-scale, and modulation-scale characteristics of environmental noise. Preferably, the following three feature channels are used: 1. The first channel focuses on the low-frequency region and a longer time window to stably extract steady-state background noise, such as air conditioner noise, fan noise, and continuous low-frequency traffic noise; 2. The second channel focuses on mid-frequency equalization analysis to take into account both the main energy range of the environment and the main operating range of the test signal; 3. The third channel focuses on the high-frequency region and a shorter time window to capture transient disturbances, such as collision sounds, the start of a speech, and short-term events like door knocking.

[0034] The three channels output the corresponding IMFCC feature tensor, frequency band energy distribution, and time modulation features, respectively. These are then weighted and integrated through a feature fusion layer to form a three-dimensional noise feature model for environmental noise. This three-dimensional noise feature model is not only used for dynamic gain compensation at the current moment but also fed back to the noise database to update the noise fingerprint classification boundaries for different scenarios in subsequent iterations. Therefore, by capturing multi-scale noise features through parallel multi-resolution channels, the three-dimensional noise feature model has higher stability, discriminative power, and generalization ability.

[0035] Here, dynamic spectrum enhancement technology is used to adaptively adjust the gain compensation amount of the test audio based on the real-time noise fingerprint. The goal is to ensure that the actual perceived signal strength is at least higher than the noise masking threshold than the target signal-to-noise ratio. Dynamic spectral enhancement calculates the signal-to-noise ratio for each 1 / 3 octave band in the frequency domain, independently calculates the required gain coefficient for each band based on the psychoacoustic masking model, and then generates a compensated time-domain test audio signal through inverse Fourier transform.

[0036] The gain compensation is the smaller of the minimum gain required to overcome the masking and the maximum safe gain allowed by the device, and is not less than 0, for example, calculated according to the following formula: in, For frequency The noise masking threshold on To test the sound pressure level of the audio signal, The target signal-to-noise ratio is (e.g., 6dB). The maximum safety gain is calculated (not exceeding the device's maximum output capability minus a 3 dB margin, which can be determined based on the device model). The noise masking threshold is calculated by the cloud server for each test frequency (e.g., 250 Hz, 500 Hz, 1 kHz, 2 kHz, 4 kHz, 8 kHz) based on the critical band masking model in ISO 28961.

[0037] S3: Obtain the reference hearing level corresponding to the current test frequency according to the test protocol; use the compensation parameters obtained in step S1 to convert the reference hearing level into a reference physical sound pressure level, and use the dynamic gain compensation amount obtained in step S2 to compensate the reference physical sound pressure level to obtain the compensated target physical sound pressure level.

[0038] In this embodiment of the invention, the target physical playback sound level suitable for the current device, frequency, and environmental conditions is obtained by combining device calibration parameters and environmental compensation. This is then mapped to the playback gain parameters of the terminal device's audio engine. Specifically, the reference hearing level, i.e., logical decibels (LdB), is nonlinearly mapped using the hyperbolic tangent mapping function according to the following formula to obtain the reference sound pressure level, i.e., physical decibels (PdB). Then, the gain compensation amount (ΔdB) calculated in real time in step S2 is used for compensation to obtain the compensated target sound pressure level (PdB_target).

[0039] In one embodiment of the present invention, a hyperbolic tangent mapping function is used to adapt auditory nonlinearity to achieve the conversion from logical decibels (LdB) to physical decibels (PdB): Where T represents the nonlinear transition threshold (dB); k represents the human ear sensitivity adjustment factor; and a, b, and c are the device physical response calibration coefficients (obtained in step S1). A piecewise hyperbolic tangent mapping function is used, with tanh⁻¹ enhancing resolution in the low sound intensity region (e.g., increasing accuracy by 0.1dB), precisely matching the logarithmic sensitivity characteristics of the human ear. When the logical decibel LdB ≥ T, linear mapping is automatically switched to avoid the saturation distortion of traditional logarithmic compression in the high-pressure region (>90dB). That is, the human ear is most sensitive to medium-intensity sounds, with saturation response at extremely high / low sound pressure levels. The hyperbolic tangent function tanh(x) also has a linear region in the center and saturation regions at both ends, perfectly simulating the nonlinear compression characteristics of hearing, avoiding extreme sound pressure levels, protecting hearing, while maintaining fine resolution at medium intensity.

[0040] Parameters a, b, and c are the device physical response calibration coefficients, which can calibrate the frequency response curves of different microphones / speakers. Based on these parameters, the conversion from logical decibels to physical decibels can be achieved, enabling the same logical value to be adapted to the physical output of multiple devices, thus improving the hardware versatility.

[0041] S4: The WebAudio application calculates the gain value based on the target sound pressure level and the device's physical response calibration coefficient (acoustic characteristics), which is used to test sound playback.

[0042] The WebAudio application locates the frequency response curve of the terminal device based on the current test frequency f (finding the device's physical response calibration coefficients at that frequency, representing the device's actual sensitivity). Upon receiving the target sound pressure level (PdB_target), the WebAudio application calculates the gain value = f(PdB, device sensitivity) using conventional techniques, accurately mapping the target sound pressure level to the WebAudio gain value (a coefficient ranging from 0 to 1), which is then used to control the output of the digital-to-analog converter (DAC).

[0043] Before each test audio playback, the cloud server calculates the dynamic gain compensation amount (ΔdB) based on real-time noise fingerprinting, combines it with the reference sound pressure level to obtain the compensated target sound pressure level (PdB), and sends it to the terminal device. The terminal device's WebAudio application queries the frequency response curve of the terminal device based on the current test frequency f (which has already been sent from the cloud server to the terminal device during the initialization phase), converts the target sound pressure level (PdB) into a gain value (0~1), sets the pure tone audiometry standard waveform parameters (time history / envelope, etc.) at the current frequency, and then calls the .start() method to achieve precise masking compensation. In this way, the audio amplitude and gain value can be dynamically adjusted through the "logical decibel-physical decibel mapping algorithm" to ensure standardized sound intensity output from different terminal devices, thereby solving the accuracy problem caused by hardware differences and achieving cross-device testing consistency.

[0044] S5: Play a pure tone test tone through the WebAudio application on the terminal device, and record the playback start timestamp.

[0045] S6: Use Automatic Speech Recognition (ASR) technology to collect audio during the test and record timestamps.

[0046] Intelligent guidance is achieved through a multimodal perception-decision-execution closed loop. The front-facing camera tracks eye focus in real time, and triggers guidance when the gaze lingers on the interface for more than 5 seconds without any operation; the pressure sensor simultaneously monitors the touch screen force, and if the touch screen force is less than the preset pressure value, it is judged as operation hesitation.

[0047] The digital human synchronously generates lip-syncing guidance animations, while enhancing the sense of speech direction and reducing environmental interference through a sound field focusing algorithm.

[0048] S7: Performs streaming speech recognition on the sound signal to obtain an effective response, and uses Transformer time-frequency masking to separate the signal, outputs a confidence score, and combines the effective response and confidence score to determine the subject's hearing level.

[0049] Multi-channel IMFCC extracts multi-scale features of environmental noise in parallel to construct an initial three-dimensional noise feature model. This model is then used as prior knowledge and input into the Transformer network. In this way, when calculating the time-frequency mask, the Transformer network incorporates these prior noise features, thereby more accurately separating the clean test tone and user speech components from the mixed signal.

[0050] The ASR module employs a streaming speech recognition engine for real-time audio reception and feedback, enabling automatic speech recognition. In one embodiment of the invention, a dual mechanism of silence detection triggering and noise immunity modeling is used to ensure the generation of a highly robust ASR module. This achieves deep coupling between speech recognition and the hearing threshold decision-making closed loop, facilitating reliable response acquisition in the hearing threshold assessment process. Specifically, this includes: 1. Integrate test tone triggering, silence detection, frequency band enhancement, template matching, or recognition engine output with the hearing threshold task flow; 2. Use Transformer time-frequency masks to extract more reliable test-related components from the mixed sound, rather than simply performing general speech denoising; 3. The identification results and confidence scores will be used together as the basis for decisions on "whether to accept this response" and "whether to retest the current frequency point"; 4. Enable the ASR module to directly participate in threshold updates, rather than just acting as an independent speech recognition module.

[0051] Meanwhile, because the mixed signal picked up by the pickup module includes test tone, ambient noise, and echo, the Transformer's self-attention mechanism is used to calculate the time-frequency mask (other algorithms can also be used) to separate the pure test tone component from the mixed signal, calculate the signal-to-noise ratio (SNR) of the actual received test tone, and output a confidence score. The confidence score output by the Transformer is defined as the SNR mapping value of the ratio of the separated test tone energy to the residual noise energy, ranging from 0 to 1. The preset confidence threshold is 0.65. When the score is ≥0.65, it is considered 'high', and the threshold can be updated; when the score is <0.65, it is considered 'low', and a retest of the current frequency point is triggered.

[0052] Based on the valid response and the confidence level of the Transformer output, if the confidence level is high, the logical decibel level is updated using an adaptive stepping method (e.g., if the user feedback indicates hearing, the first preset value is decreased by 10dB; if not heard, the second preset value is increased by 5dB). If the confidence level is low, the current frequency point is retested (up to 2 times). If the threshold determination for the current frequency point is completed, the test frequency is moved to the next test frequency. This ensures the integrity of the main frequency band of the speech and prevents the loss of effective components of the test tone due to filtering.

[0053] The cloud server continuously updates the noise database through a closed-loop self-iterative mechanism, and the Transformer time-frequency masking algorithm ensures the integrity of the main frequency band of speech, providing a clinical-grade, precise acoustic environment adaptation solution for hearing-impaired patients.

[0054] Based on high confidence and a valid response (user feedback of hearing), and according to the WHO 2021 hearing loss classification criteria (using the mean pure tone thresholds of 0.5, 1, 2, and 4 kHz as the standard), the subject's hearing level was determined: Normal (≤20 dB HL) Mild (20–35 dB HL) Moderate (35–50 dB HL) Moderate to severe (50–65 dB HL) Severe (65–80 dB HL) Extremely severe (≥80 dB HL).

[0055] Then, the audiogram is displayed on the terminal device in large font and graphical format, along with a voice explanation, such as, "You have moderate hearing loss in your left ear at 4 kHz. It is recommended to avoid noisy environments and consider further professional examination." Personalized intervention suggestions are also provided. Mild hearing loss: Regular check-ups and use of hearing aids (such as amplifying ambient sounds) are recommended. Moderate or worse: Referral to an ENT specialist is recommended, and a list of nearby medical facilities (based on GPS) should be provided. If asymmetry of ≥15 dB (at the same frequency) is detected in both ears, it is noted that "it is recommended to seek medical attention as soon as possible to rule out conductive or retrocochlear lesions".

[0056] S8: Collect data and environmental noise from all test sessions, update the 3D noise feature model, and continuously update the noise database and the acoustic characteristic database of the terminal devices.

[0057] Each test session collects data (anonymized), including environmental noise characteristics (3D noise feature model), user feedback and final thresholds, Transformer mask verification results (signal-to-noise ratio, confidence level), device model and calibration parameter usage, timestamps, etc. The following offline processes are performed periodically (e.g., weekly) and the data in the cloud server is updated: 1) Retraining the Transformer time-frequency mask neural network with new data; 2) Optimizing the filter parameters of the multi-channel IMFCC; 3) Updating the compensation rules for dynamic spectrum enhancement (e.g., masking coefficients); 4) Expanding the noise database (e.g., adding "TV background noise," "kitchen noise," etc.); 5) Optimizing the parameters (T, k, a, b, c) of the piecewise hyperbolic tangent mapping.

[0058] The updated model parameters will be sent to the terminal device via OTA in the next test initialization step (step S1).

[0059] To ensure millisecond-level response times for cloud server services, the hot updates (inference) and cold updates (retraining) of the 3D noise feature model in this embodiment of the invention are separated. Real-time inference uses the current version of the model parameters; while offline retraining utilizes accumulated big data, which can continuously improve the environmental adaptability of the 3D noise feature model without interrupting service.

[0060] The actual assessment results are provided below: Table 1 Experimental conditions for subjects Table 2. Subject completion rate data Table 3. Results of loudness assessment by the subjects Table 4 Frequency Evaluation Results Table 5 Comparison data before and after compensation Test results showed that 28 out of 30 participants successfully completed all tests, achieving an overall completion rate of 93.3%, with an average test duration of 11.8 minutes. In the loudness assessment phase, the average threshold for the left ear was 41.2 dB, and for the right ear, it was 39.8 dB. In the frequency assessment phase, the proportion of hearing impairment increased gradually with increasing test frequency, with significantly higher abnormalities at 4000 Hz and 8000 Hz compared to lower frequencies, consistent with common characteristics of high-frequency hearing threshold elevation in elderly participants. Further comparisons showed that without device compensation and dynamic compensation, the average threshold deviation between different terminal devices reached 9.6 dB; after enabling compensation, this deviation decreased to 3.1 dB. In a typical home environment, after enabling dynamic compensation and reliability control, the repeat test deviation decreased from 8.2 dB to 3.8 dB, and the retest rate decreased from 18.5% to 9.7%. The above results demonstrate that the device difference compensation, environment adaptation, dynamic gain control, and response reliability control mechanisms provided by this invention can effectively improve the consistency, stability, and usability of hearing threshold assessment in non-professional scenarios.

[0061] In summary, based on the nonlinear perception characteristics of the human auditory system, the noise masking effect, and the common hearing loss characteristics of the elderly, this invention constructs a dynamic parameter mapping and compensation mechanism for terminal device-based hearing threshold assessment: the baseline hearing level dB HL (set value) specified in the test protocol is converted into a baseline sound pressure level dB SPL by querying a dynamic, device-specific lookup mapping table. The sum of the baseline sound pressure level dB SPL and the real-time calculated dynamic gain compensation amount (ΔdB) yields the compensated target sound pressure level, which is then used for playback.

[0062] This invention utilizes noise fingerprints extracted from multi-channel IMFCC as input for dynamic spectrum enhancement. The compensation test tone output by dynamic spectrum enhancement is verified by a Transformer mask. The verification result drives adaptive stepping method decision-making, forming real-time perception and compensation. Decisions are made only after verification, achieving real-time forward collaboration. Moreover, when insufficient signal-to-noise ratio is detected by Transformer verification, the gain parameters of dynamic spectrum enhancement are adjusted in real time. After the adaptive stepping method changes the logical decibel, dynamic spectrum enhancement recalculates the mask compensation, thus forming a mutually corrective feedback mechanism to achieve real-time feedback collaboration. In addition, all test data (noise characteristics, confidence levels, user feedback) are aggregated into a closed-loop self-iterative module to update IMFCC parameters, dynamic spectrum enhancement rules, Transformer model weights, and the noise database offline, achieving long-term iterative evolution. Such collaborative effects enable a clinical-grade precise acoustic environment even in a home setting, ensuring the effectiveness and accuracy of hearing threshold assessment.

[0063] Through the above steps, the local terminal device uploads noise data to a cloud server using wireless communication methods such as 5G. The cloud server's algorithm dynamically analyzes the noise intensity and frequency characteristics, automatically adjusting the sound intensity compensation parameters of the test audio to ensure that the test signal can still be effectively transmitted in complex environments. Because this invention improves the accuracy of assessment through adaptive environmental noise compensation and acoustic calibration of the terminal device, and utilizes intelligent guidance to achieve elderly-friendly interaction, enhancing compliance and completeness, it solves the problems of environmental noise interference, large equipment differences, operational difficulties for the elderly, and reliance on professional guidance in existing technologies. Therefore, it can be widely applied to adaptive threshold assessment in hearing screening scenarios at home, in communities, and in primary healthcare institutions.

[0064] Second Embodiment The second embodiment of the present invention also provides an intelligent hearing threshold assessment system, including a terminal device and a cloud server, for implementing the intelligent hearing threshold assessment method described in the first embodiment. The terminal device includes, but is not limited to, smartphones, tablets, or headphones, and has a built-in or external environmental noise sensor, speaker, microphone, display screen, and communication module. The cloud server includes a device acoustic characteristic database, a noise database, a model inference engine, and a parameter update module. The terminal device and the cloud server interact via a wireless communication network: the terminal device is responsible for collecting environmental noise, reporting the device model, playing test audio, picking up user responses, and displaying assessment results; the cloud server is responsible for querying the mapping table based on the device model and issuing compensation parameters, calculating dynamic gain compensation based on real-time noise, performing streaming speech recognition and Transformer time-frequency masking algorithms, outputting confidence scores, and determining the hearing level. The specific workflow of this intelligent hearing threshold assessment system corresponds to the steps of the intelligent hearing threshold assessment method described in the first embodiment, and will not be repeated here.

[0065] The intelligent hearing threshold assessment method and system provided by this invention have been described in detail above. Any obvious modifications made by those skilled in the art without departing from the essence of this invention will constitute an infringement of the patent rights of this invention and will incur corresponding legal liability.

Claims

1. A method for intelligent hearing threshold assessment, characterized in that... Includes the following steps: S1: During the initialization phase, compensation parameters are pushed according to the acoustic characteristics of the terminal device, and the compensation parameters are used to compensate the playback sound of the terminal device in real time to ensure that different terminal devices output the same physical sound pressure level. S2: During the testing phase, noise is collected in real time before the test sound is played, which is used to calculate the noise masking threshold and dynamic gain compensation amount. S3: Obtain the reference hearing level corresponding to the current test frequency according to the test protocol; use the compensation parameters obtained in step S1 to convert the reference hearing level into a reference sound pressure level, and use the dynamic gain compensation amount obtained in step S2 to compensate the reference sound pressure level to obtain the compensated target sound pressure level. S4: The gain value is calculated based on the target sound pressure level and acoustic characteristics and is used to test sound playback; S5: Play a pure tone test tone and record the playback start timestamp at the same time; S6: Collect audio during the test and record timestamps; S7: Perform streaming speech recognition on the sound signal to obtain an effective response, and simultaneously perform signal separation to output a confidence score. Combine the effective response and the confidence score to determine the subject's hearing level.

2. The intelligent hearing threshold assessment method as described in claim 1, characterized in that: In step S1, the compensation parameters are obtained by querying the mapping table and matching the compensation parameters {a, b, c, frequency response curve}, where a represents the sensitivity coefficient of the sound played by the terminal device; b represents the frequency response slope compensation coefficient; and c represents the reference offset constant. The query mapping table contains the device model and the frequency response characteristic parameters that match each device model.

3. The intelligent hearing threshold assessment method as described in claim 1, characterized in that: In step S7, the Transformer time-frequency masking algorithm is used to separate noise from potential speech / test signals to ensure that noise feature extraction is not contaminated by sudden interference.

4. The intelligent hearing threshold assessment method as described in claim 3, characterized in that: A multi-channel IMFCC feature extraction algorithm is used to process multiple feature streams with different time-frequency resolutions or different frequency bands of interest in parallel, and to perform millisecond-level spectrum tracking of noise intensity and frequency characteristics. A three-dimensional noise feature model including intensity, frequency, and time modulation characteristics is constructed to effectively distinguish between steady-state and transient noise.

5. The intelligent hearing threshold assessment method as described in claim 4, characterized in that: The term "multi-channel" refers to multiple feature channels constructed using a multi-resolution parallel analysis method to address the different time scale, frequency scale, and modulation scale characteristics of environmental noise.

6. The intelligent hearing threshold assessment method as described in claim 5, characterized in that... It also includes step S8: collecting data and environmental noise from all test sessions, updating the three-dimensional noise feature model, and continuously updating the noise database and the acoustic characteristic database of the terminal device.

7. The intelligent hearing threshold assessment method according to any one of claims 1 to 6, characterized in that: Dynamic spectrum enhancement technology is used to adaptively adjust the gain compensation amount of the test audio based on real-time noise fingerprint, so that the actual perceived signal intensity is at least higher than the target signal-to-noise ratio than the noise masking threshold. The gain compensation amount is the smaller value between the minimum gain required to overcome the masking and the maximum safe gain allowed by the device, and is not less than 0.

8. The intelligent hearing threshold assessment method as described in claim 7, characterized in that: In step S3, the reference hearing level specified in the test protocol is nonlinearly mapped to obtain the reference sound pressure level; then, the gain compensation amount calculated in real time in step S2 is used for compensation to obtain the compensated target sound pressure level.

9. The intelligent hearing threshold assessment method as described in claim 8, characterized in that: The hyperbolic tangent mapping function is used to adapt to auditory nonlinearity and convert the reference hearing level to the reference sound pressure level.

10. An intelligent hearing threshold assessment system, characterized in that... include: Terminal devices are used to collect environmental noise, play test audio, pick up user responses, and display evaluation results; A cloud server is used to implement the steps in the intelligent hearing threshold assessment method according to any one of claims 1 to 9; The terminal device and the cloud server interact with each other via a wireless communication network. The cloud server sends compensation parameters based on a lookup mapping table of the device model reported by the terminal device; calculates dynamic gain compensation based on the environmental noise collected in real time by the terminal device; executes streaming speech recognition and Transformer time-frequency masking algorithms, outputs a confidence score, and determines the hearing level. The terminal device then performs sound pressure level compensation on the test audio based on the compensation parameters and dynamic gain compensation sent by the cloud server before playing it.

Images