- Research Article
- Open Access
Online Personalization of Hearing Instruments
EURASIP Journal on Audio, Speech, and Music Processing volume 2008, Article number: 183456 (2008)
Online personalization of hearing instruments refers to learning preferred tuning parameter values from user feedback through a control wheel (or remote control), during normal operation of the hearing aid. We perform hearing aid parameter steering by applying a linear map from acoustic features to tuning parameters. We formulate personalization of the steering parameters as the maximization of an expected utility function. A sparse Bayesian approach is then investigated for its suitability to find efficient feature representations. The feasibility of our approach is demonstrated in an application to online personalization of a noise reduction algorithm. A patient trial indicates that the acoustic features chosen for learning noise control are meaningful, that environmental steering of noise reduction makes sense, and that our personalization algorithm learns proper values for tuning parameters.
Modern digital hearing aids contain advanced signal processing algorithms with many tuning parameters. These are set to values that ideally match the needs and preferences of the user. Because of the large dimensionality of the parameter space and unknown determinants of user satisfaction, the tuning procedure becomes a complex task. Some of the tuning parameters are set by the hearing aid dispenser based on the nature of the hearing loss. Other parameters may be tuned on the basis of the models for loudness perception, for example . But, not every individual user preference can be put into the hearing aid beforehand because some particularities of the user may be hard to represent into the algorithm, and the user's typical acoustic environments may be very different from the sounds that are played to the user in a clinical fitting session. Moreover, sound preferences may be changing with continued wear of a hearing aid. Thus, users sometimes return to the clinic soon after the initial fitting for further adjustment . In order to cope with the various problems for tuning parameters prior to device usage, we present in this paper a method to personalize the hearing aid algorithm during usage to actual user preferences.
We consider the personalization problem as linear regression from acoustic features to tuning parameters, and formulate learning in this model as the maximization of an expected utility function. An online learning algorithm is then presented that is able to learn preferred parameter values from control operations of a user during usage. Furthermore, when a patient leaves the clinic with a fitted hearing aid, it is not completely known which features are relevant for explaining the patient's preference. Taking "just every interesting feature" into account may lead to high-dimensional feature vectors, containing irrelevant and redundant features that make online computations expensive and hinder generalization of the model. Irrelevant features do not contribute to predicting the output, whereas redundancy refers to features that are correlated with other features which do not contribute to the output when the correlated features are also present. We therefore study a Bayesian feature selection scheme that can learn a sparse and well-generalizing model for observed preference data. The behavior of the Bayesian feature selection scheme is validated with synthetic data, and we conclude that this scheme is suitable for the analysis of hearing aid preference data. An analysis of preference data from a listening test reveals a relevant set of acoustic features for personalized noise reduction.
Based on these features, a learning noise control algorithm was implemented on an experimental hearing aid. In a patient trial, 10 hearing impaired subjects were asked to use the experimental hearing aid in their daily life for six weeks. The noise reduction preferences showed quite some variation over subjects, and most of the subjects learned a preference that showed a significant dependency on acoustic environment. In a post hoc sound quality analysis, each patient had to choose between the learned hearing aid settings and a (reasonable) default setting of the instrument. In this blind laboratory test, 80% of the subjects preferred the learned settings.
This paper is organized as follows. In Section 2, the model for hearing aid personalization is described, including algorithms for both offline and online training of tuning parameters. In Section 3, the Bayesian feature selection algorithm is quickly reviewed along with two fast heuristic feature selection methods. In addition, the methods are validated experimentally. In Section 4, we analyze a dataset with noise reduction preferences from an offline data collection experiment in order to obtain a reduced set of features for online usage. A clinical trial to validate our online personalization model is presented in Section 5. Section 6 discusses the experimental results, and we conclude in Section 7.
2. A Model for Hearing Aid Personalization
Consider a hearing aid (HA) algorithm , where and are the input and output signals, respectively, and is a vector of tuning parameters, such as time constants and thresholds. HA algorithms are by design compact in order to save energy consumption. Still, we want that performs well for all environmental conditions. As a result, good values for the tuning parameters are often dependent on the environmental context, like being in a car, a restaurant setting, or at the office. This will require a tuning vector that varies with time (as well as context). Many hearing aids are equipped with a so-called control wheel (CW), which is often used by the patient to adjust the output volume (cf. Figure 1). Online user control of a tuning parameter does not need to be limited to the volume parameter. In principle, the value of any component from the tuning parameter vector could be controlled through manipulation of the CW. In this paper, we will denote by a scalar tuning parameter that is manually controlled through the CW.
2.1. Learning from Explicit Consent
An important issue concerns how and when to collect training data. When a user is not busy manipulating the CW, we have no information about his satisfaction level. After all, the patient might not be wearing the instrument. When a patient starts with a CW manipulation, it seems reasonable to assume that he is not happy with the performance of his instrument. This moment is tagged as a dissent moment. Right after the patient has finished turning the CW, we assume that the patient is satisfied with the new setting. This moment is identified as a consent moment. Dissent and consent moments identify situations for collecting training data that relate to low and high satisfaction levels. In this paper, we will only learn from consent moments.
Consider the system flow diagram of Figure 2. The tuning parameter value is determined by two terms. The user can manipulate the value of directly through turning a control wheel. The contribution to from the CW is called (for "manual"). We are interested however in learning separate settings for under different environment conditions. For this purpose, we use an EnVironment Coder (EVC) that computes a -dimensional feature vector based on the input signal . The feature vector may consist of acoustic descriptors like input power level and speech probability. We then combine the environmental features linearly through , and add this term to the manual control term, yielding
We will tune the "environmental steering" parameters based on data obtained at consent moments. We need to be careful with respect to the index notation. Assume that the th consent moment is detected at ; that is, the value of the feature vector at the th consent moment is given by . Since our updates only take place right after detecting the consent moments, it is useful to define a new time series as
as well as similar definitions for converting to . The new sequence, indexed by rather than , only selects samples at consent moments from the original time series. Note the difference between and . The latter () refers to one sample (e.g., millisecond) after the consent moment , whereas was measured at the th consent moment, which may be hours after .
Again, patients are instructed to use the control wheel to tune their hearing instrument at any time to their liking. Just seconds before consent moment , the user experiences an output that is based on a tuning parameter . Notation refers to the value for prior to the th user action. Since is considered small with respect to typical periods between consent times and since we assume that features are determined at a time scale that is relatively large with respect to , we make the additional assumption that . Hence, adjusted settings at time are found as
The values of the tuning parameter and the features are recorded at all registered consent moments, leading to the preference dataset
We assume that the user generates tuning parameter values at consent times via adjustments , according to a preferred steering function
where are the steering parameter values that are preferred by the user, and are the preferred (environment-dependent) tuning parameter values. Due to dexterity issues, inherent uncertainty on the patient's part, and other disturbing influences, the adjustment that is provided by the user will contain noise. We model this as an additive white Gaussian "adjustment noise" contribution to the "ideal adjustment" (and with we mean a variable that is distributed as a normal distribution with mean and covariance matrix ). Hence, our model for the user adjustment is
Consequently, our preference data is generated as
Since the preferred steering vector is unknown and we want to predict future values for the tuning parameter , we introduce stochastic variables and and propose the following probabilistic generative model for the preference data:
According to (8), the probability of observing variable is conditionally Gaussian:
We now postulate that minimization of the expected adjustment noise will lead to increased user satisfaction since predicted values for the tuning parameter variable will be more reflecting the desired values. Hence, we define a utility function for the personalization problem:
where steering parameters are now also used as utility parameters. We find personalized tuning parameters by setting them to the value that maximizes the expected utility for the user:
The maximum expected utility is reached when we set
where is the posterior mean of the utility parameters:
The goal is therefore to infer the posterior over the utility parameters given a preference dataset . During online processing, we find the optimal tuning parameters as
The value for can be learned either offline or online. In the latter case, we will make recursive estimates of , and apply those instead of .
Our personalization method is shown schematically in Figure 3, where we represent the uncertainty in the user action as a behavioral model that links utilities to actions by applying an exponentiation to the utilities.
2.3. Offline Training
If we perform offline training, we let the patient walk around with the HA (or present acoustic signals in a clinical setting), and let him manipulate the control wheel to his liking in order to collect an offline dataset as in (4). To emphasize the time-invariant nature of in an offline setting, we will omit the index from . Our goal is then to infer the posterior over the utility parameters given dataset :
where prior is defined as
and the likelihood term equals
Then, the maximum a posteriori solution for is
and coincides with the MMSE solution. Here, we defined and the -dimensional feature matrix . By choosing a different prior , one may, for example, emphasize sparsity in the utility parameters. In Section 3, we will evaluate a method for offline regression that uses a marginal prior that is more peaked than a Gaussian one, and hence it performs sound feature selection and fitting of utility parameters at the same time. Such an offline feature selection stage is not strictly necessary, but it can make the consecutive online learning stage in the field more (computationally) efficient.
2.4. Online Training
During online training, the parameters are updated after every consent moment . The issue is then how to update on the basis of the new data . We will now present a recursive algorithm for computing the optimal steering vector , that is, enabling online updating of . We leave open the possibility that user preferences change over time, and allow the steering vector to "drift" with some white Gaussian (state) noise . Hence, we define observation vector and state vector as stochastic variables with conditional probabilities and , respectively. In addition, we specify a prior distribution . This leads to the following state space model for online preference data:
We can recursively estimate the posterior probability of given new user feedback :
according to the Kalman filter :
where and are (time-varying) state and observation noise variances. The rate of learning in this algorithm depends on these noise variances. Online estimates of the noise variances can be made by the Jazwinski method  or by using recursive EM. The state noise can become high when a transition to a new dynamic regime is experienced. The observation noise measures the inconsistency in the user response. The more consistently the user operates the control wheel, the less the estimated observation noise and the higher the learning rate will be.
In summary, after detecting the th consent, we update according to
2.5. Leaving the User in Control
As mentioned before, we use the posterior mean to update steering vector with a factor of . By itself, an update would cause a shift in the perceived value for tuning parameter . In order to compensate for this undesired effect, the value of the control wheel register is decreased by the same amount. The complete online algorithm (excluding Kalman intricacies) is shown in Figure 4. In our algorithm, we update the posterior over the steering parameters immediately after each user control action, but the effect of the updating becomes clear to the user only when he enters a different environment (which will lead to very different acoustical features ). Further, the "optimal" environmental steering (i.e., without the residual ) is applied to the user at a much larger time scale. This ensures that the learning part of the algorithm (lines (5)–(7)) leads to proper parameter updates, whereas the steering part (line (3)) does not suffer from sudden changes in the perceived sounds due to a parameter update. We say that "the user remains in control" of the steering at all times.
By maximizing the expected utility function in (10), we focus purely on user consent; we consider a new user action as "just" the generation of a new target value . We have not (yet) modeled the fact that the user will react on updated settings for , for example, because these settings lead to unwanted distortions or invalid predictions for in acoustic environments for which no consent was given. The assumption is that any induced distortions will lead to additional user feedback, which can be handled in the same manner as before.
Note that by avoiding a sense of being out of control, we effectively make the perceived distortion part of the optimization strategy. In general, a more elaborate model would fully close the loop between hearing aid and user by taking expected future user actions into account. We could then maximize an expected "closed-loop" utility function , where is shorthand for the earlier utility function of (10), utility term expresses other perceived distortions, and utility term reflects the cost of making (too many) future adjustments.
2.6. Example: A Simulated Learning Volume Control
We performed a simulation of a learning volume control (LVC), where we made illustrative online regression of broadband gain (volume = ) at input power level (log of smoothed RMS value of the input signal = ). As input, we used a music excerpt that was preprocessed to give one-dimensional log-RMS feature values. This was fed to a simulated user who was supposed to have a (one-dimensional) preferred steering vector . During the simulation, noisy corrections were fed back from the user to the LVC in order to make the estimate resemble the preferred steering vector . We simulated a user who has time-varying preferences. The preferred value changed throughout the input that was played to the user, according to consecutive preference modes , and . With , we mean the preferred value during mode . A mode refers to a preferred value during a consecutive set of time samples when playing the signal. Further, feature values are negative in this example. Therefore a negative value of leads to an effective amplification, and vice versa for positive .
Moreover, the artificial user experiences a threshold on his annoyance, which will determine if he will make an actual adjustment. When the updated value comes close to the desired value at the corresponding time, the user stops making adjustments. Here we predefined a threshold on the difference to quantify "closeness." In the simulation, the threshold was put to 0.02; this will lead to many user adjustments for the nonlearning volume control situation. Increasing this threshold value will lead to less difference in the amount of user adjustments between learned and nonlearned cases. When the difference between updated and desired values exceeds the threshold, the user will feed back a correction value proportional to the difference , to which Gaussian adjustment noise is added. The variance of the noise changed throughout the simulation according to a set of "consistency modes." Finally, we omitted the discount operation in this example since we merely use this example to illustrate the behavior of inconsistent users with changing preferences.
We analyzed the behavior when the LVC was part of the loop, and compared this to the situation without an LVC. In the latter case, user preferences are not captured in updated values for , and the user annoyance (as measured by the number of user actions) will be high throughout the simulation. In Figure 5(a), we show the (smoothed) log-RMS value of the desired output signal in blue. The desired output signal is computed as , where is the smoothed log-RMS value of input signal , and is some fixed function that determines how the predicted hearing aid parameter is used to modify the incoming sound. The log-RMS of the realized output signal is plotted in red. The value for is fixed to zero in this simulation (see Figure 5(b)). Any noise in the adjustments will be picked up in the output unless the value for happens to be close to the fixed value . We see in Figure 5 that the red curve resembles a noisy version of the blue (target) curve, but this comes at the expense of many user actions. Any nonzero value in Figure 5(c) reflects one noisy user adjustment. When we compare this to Figure 6, we see that by using an LVC we achieve a less noisy output realization (see Figure 6(a)) and proper tracking of the four preference modes (see Figure 6(b)) by a relatively small number of user adjustments (see Figure 6(c)). Note that the horizontal axis in the former figures is in seconds, demonstrating that this simulation is in no way realistic of real-world personalization. It is included to illustrate that in a highly artificial setup an LVC may diminish the number of adjustments when the noise in the adjustments is high and the user preference changes with time. We study the real-world benefits of an algorithm for learning control in Section 5.
3. Acoustic Feature Selection
We now turn to the problem of finding a relevant (and nonredundant) set of acoustic features in an offline setting. Since user preferences are expected to change mainly over long-term usage, the coefficients are considered stationary for a certain data collection experiment. In this section, three methods for sparse linear regression are reviewed that aim to select the most relevant input features in a set of precollected preference data. The first method, Bayesian backfitting, has a great reputation for accurately pruning large-dimensional feature vectors, but it is computationally demanding . We also present two fast heuristic feature selection methods, namely, forward selection and backward elimination. In this section, both of the Bayesian and heuristic feature selection methods are quickly reviewed, and experimental evaluation results are presented. To emphasize the offline nature, we will index samples with rather than with or in the remainder of this section, or drop the index when the context is clear.
3.1. Bayesian Backfitting Regression
Backfitting  is a method for estimating the coefficients of linear models of the form
Backfitting decomposes the statistical estimation problem into individual estimation problems by creating "hidden targets" for each term (see Figure 7). It decouples the inference in each dimension, and can be solved with an efficient expectation-maximization (EM) algorithm that avoids matrix inversion. This can be a very lucrative option if the input dimensionality is large. A probabilistic version of backfitting has been derived in , and in addition it is possible to assign prior probabilities to the coefficients . For instance, if we choose
as (conditional) priors for and , then it can be shown  that the marginal prior over the coefficients is a multidimensional Student's -distribution, which places most of its probability mass along the axial ridges of the space. At these ridges, the magnitude of only one of the parameters is large; hence this choice of prior tends to select only a few relevant features. Because of this so-called automatic relevance determination (ARD) mechanism, irrelevant or redundant components will have a posterior mean ; so the posterior distribution over the corresponding coefficient will be narrow around zero. Hence, the coefficients that correspond to irrelevant or redundant input features become zero. Effectively, Bayesian backfitting accomplishes feature selection and coefficient optimization in the same inference framework.
We have implemented the Bayesian backfitting procedure by the variational EM algorithm [5, 8], which is a generalization of the maximum likelihood-based EM method. The complexity of the full variational EM algorithm is linear in the input dimensionality (but scales less favorably with sample size). Variational Bayesian (VB) backfitting is a fully automatic regression and feature selection method, where the only remaining hyperparameters are the initial values for the noise variances and the convergence criteria for the variational EM loop.
3.2. Fast Heuristic Feature Selection
For comparison, we present two fast greedy heuristic feature selection algorithms specifically tailored for the task of linear regression. The algorithms apply (1) forward selection (FW) and (3) backward elimination (BW), which are known to be computationally attractive strategies that are robust against overfitting . Forward selection repetitively expands a set of features by always adding the most promising unused feature. Starting from an empty set, features are added one at a time. Once, selected features have been never removed. Backward elimination employs the reverse strategy of FW. Starting from the complete set of features, it generates an ordering at each time taking out the least promising feature. In our implementation, both algorithms apply the following general procedure.
For all features and outputs, subtract the mean and scale to unit variance. Remove features without variance. Precalculate second-order statistics on full data.
Repeat 10 times.
(a)Split dataset: randomly take out of the samples for validation. The statistics of the remaining 90% are used to generate the ranking.
(b)Heuristically rank the features (see below).
(c)Evaluate the ranking to find the number of features that minimizes the validation error.
From all 10 values (found at 2c), select the median . Then, for all rankings, count the occurrences of a feature in the top to select the most popular features, and finally optimize their weights on the full dataset.
The difference between the two algorithms lies in the ranking strategy used at step 2b. To identify the most promising feature, FW investigates each (unused) feature, directly calculating training errors using (B.5) of Appendix B. In principle, the procedure can provide a complete ordering of all features. The complexity, however, is dominated by the largest sets; so needlessly generating them is rather inefficient. FW therefore stops the search early when the minimal validation error has not decreased for at least 10 runs. To identify the least promising feature, our BW algorithm investigates each feature still being a part of the set and removes the one that provides the largest reduction (or smallest increase) of the criterion in (B.5). Since BW spends most of the time at the start, when the feature set is still large, not much can be gained using an early stopping criterion. Hence, in contrast to FW, BW always generates a complete ordering of all features. Much of the computational efficiency in the benchmark feature selection methods comes from a custom-designed precomputation of data statistics (see Appendix B).
3.3. Feature Selection Experiments
We compared the Bayesian feature selection method to the benchmark methods with respect to the ability to detect irrelevant and redundant features. For this purpose, we generated artificial regression data according to the procedure outlined in Appendix A. We denote the total number of features in a dataset by , and the number of irrelevant features by . The number of redundant features is , and the number of relevant features is . The aim in the next two experiments is to find a value for (the number of selected features) that is equal to the number of relevant features in the data.
3.3.1. Detecting Irrelevant Features
In a first experiment, the number of relevant features is and . Specifically, the first and the last five input features were irrelevant for predicting the output, and all other features were relevant. We varied the number of samples as , and studied two different dimensionalities . We repeated 10 runs of each feature selection experiment (each time with a new draw of the data), and trained both Bayesian and heuristic feature selection methods on the data. The Bayesian method was trained for 200.000 cycles at maximum or when the likelihood improved less than 1e-4 per iteration, and we computed the classification error for each of the three methods. A misclassification is a feature that is classified as relevant by the feature selection procedure, whereas it is irrelevant or redundant according to the data generation procedure, and v.v. The classification error is the total number of misclassifications in 10 runs normalized by the total number of features present in 10 runs. The mean classification results over 10 repetitions (the result for is based on 5 runs) are shown in Figure 8. We see that for both 15 and 50 features and for moderate to high sample sizes (where we define moderate sample size as for and for , VB outperforms FW and performs similar to BW. For small sample sizes, FW and BW outperform VB.
3.3.2. Detecting Redundant Features
In a second experiment, we added redundant features to the data; that is, we included optional step 4 in the data generation procedure of Appendix B. The number of redundant features is , and equals the number of relevant features . In this experiment, was varied and the output SNR was fixed to 10. The role of relevant and redundant features may be interchanged, since a rotated set of relevant features may be considered by a feature selection method as more relevant than the original ones. In this case, the originals become the redundant ones. Therefore, we determined the size of the redundant subset in each run (which should equal for , resp.). In Figure 9, we plot the mean size of the redundant subset over 10 runs for different , including one-standard-deviation error bars. For moderate sample sizes, both VB and the benchmark methods detect the redundant subset (though they are biased to somewhat larger values), but accuracy of the VB estimate drops with small or large sample sizes (for explanation, see ). We conclude that VB is able to detect both irrelevant and redundant features in a reliable manner for dimensionalities up to 50 (which was the maximum dimensionality studied) and moderate sample sizes. The benchmark methods seem to be more robust to small sample problems.
4. Feature Selection in Preference Data
We implemented a hearing aid algorithm on a real-time platform, and turned the maximum amount of noise attenuation in an algorithm for spectral subtraction into an online modifiable parameter. To be precise, when performing speech enhancement based on spectral subtraction (see, e.g., ), one observes noisy speech , and assumes that speech and noise are additive and uncorrelated. Therefore, the power spectrum of the noisy signal is also additive: . In order to enhance the noisy speech, one applies a gain function in frequency bin , to compute the enhanced signal spectrum as . This requires an estimate of the power spectrum of the desired signal since, for example, the power spectral subtraction gain is computed as . If we choose the clean speech spectrum as our desired signal, an attempt is made to remove all the background noise from the signal. This is often unwanted since it leads to audible distortions and loss of environmental awareness. Therefore, one can also choose , where is a parameter that controls the remaining noise floor. The optimal setting of gain depth parameter is expected to be user- and environment-dependent. In the experiments with learning noise control, we therefore let the user personalize an environment-dependent gain depth parameter.
Six normal hearing subjects were exposed in a lab trial to an acoustic stimulus that consisted of several speech and noise snapshots picked from a database (each snapshot is typically in the order of 10 seconds), which were combined in several ratios and appended. This led to one long stream of signal/noise episodes with different types of signals and noise in different ratios. The subjects were asked to listen to this stream several times in a row and to adjust the noise reduction parameter as desired. Each time an adjustment was made, the acoustic input vector and the desired noise reduction parameter were stored. At the end of an experiment, a set of input-output pairs was obtained from which a regression model was inferred using offline training.
We postulated that two types of features are relevant for predicting noise reduction preferences. First, a feature that codes for speech intelligibility is likely to explain some of the underlying variance in the regression. We proposed three different "speech intelligibility indices:" speech probability (PS), signal-to-noise ratio (SNR), and weighted signal-to-noise ratio (WSNR). The PS feature measures the probability that speech is present in the current acoustic environment. Speech detection occurs with an attack time of 2.5 seconds and a release time of 10 seconds. These time windows refer to the period during which speech probability increases from 0 to 1 (attack), or decreases from 1 to 0 (release). PS is therefore a smoothed indicator of the probability that speech is present in the current acoustic scene, not related to the time scales (of milliseconds) at which a voice activity detector would operate. The SNR feature is an estimate of the average signal-to-noise ratio in the past couple of seconds. The WSNR feature is a signal-to-noise ratio as well, but instead of performing plain averaging of the signal-to-noise ratios in different frequency bands, we now weight each band with the so-called "band importance function"  for speech. This is a function that puts higher weight to bands where speech has usually more power. The rationale is that speech intelligibility will be more dependent on the SNR in bands where speech is prevalent. Since each of the features PS, SNR and WSNR codes for "speech presence," we expect them to be correlated.
Second, a feature that codes for perceived loudness may explain some of the underlying variance. Increasing the amount of noise reduction may influence the loudness of the sound. We proposed broadband power (Power) as a "loudness index," which is likely to be uncorrelated with the intelligibility indices. The features WSNR, SNR, and Power were computed at time scales of , and seconds, respectively. Since PS was computed at only one set of (attack and release) time scales, this led to features. The number of adjustments for each of the subjects was . This means that we are in the realm of moderate sample size and moderate dimensionality, for which VB is accurate (see Section 3.3).
We then trained VB on the six datasets. In Figure 10, we show for four of the subjects a Hinton diagram of the posterior mean values for the variance (i.e., ). Since the PS feature is determined at a different time scale than the other features, we plotted the value of that was obtained for PS on all positions of the time scale axis. Subjects 3 and 6 adjust the hearing aid parameter primarily based on feature types: Power and WSNR. Subjects 1 and 5 only used the Power feature, whereas subject 4 used all feature types (to some extent). Subject 2 data could not be fit reliably (noise variances were high for all components). No evidence was found for a particular time scale since relevant features are scattered throughout all scales. Based on these results, broadband power and weighted SNR were selected as features for a subsequent clinical trial. Results are described in the next section.
5. Hearing Aid Personalization in Practice
To investigate the relevance of the online learning model and the previously selected acoustic features, we set up a patient trial. We implemented an experimental learning noise control on a hearing aid, where we used the previously selected features for prediction of the maximum amount of attenuation in a method for spectral subtraction. During the trial, 10 hearing impaired patients were fit with these experimental hearing aids. Subjects were uninformed about the fact that it was a learning control, but only that manipulating the control would influence the amount of noise in the sound. The full trial consisted of a field trial, a first lab test halfway through the field trial, and a second lab test after the field trial. During the first fitting of the hearing instruments (just before the start of the field trial), a speech perception in noise task was given to each subject to determine the speech reception threshold in noise , that is, the SNR needed for an intelligibility score of 50%.
5.1. Lab Test 1
In the first lab test, a predefined set of acoustic stimuli in a signal-to-noise ratio range of [ dB, 10 dB] and a sound power level range of [50 dB, 80 dB] SPL was played to the subjects. SPL refers to sound pressure level (in dB) which is defined as , where is the pressure of the sound that is measured and is the sound pressure that corresponds to the hearing threshold (and no A-weighting was applied to the stimuli). The subjects were randomly divided into two test groups, A and B, in a cross-over design. Both groups started with a first training phase, and they were requested to manipulate the hearing instrument on a set of training stimuli during 10 minutes in order to make the sound more pleasant. This training phase modified the initial (default) setting of 8 dB noise reduction into more preferred one. Then, a test phase contained a placebo part and a test part. Group A started with the placebo part followed by the test part, and group B used the reversed order. In the placebo part, we played another set of sound stimuli during 5 minutes, where we started with default noise reduction settings and again requested to manipulate the instrument. In the test part of the test phase, the same stimulus as in the placebo part was played but training continued from the learned settings from the training session. Analysis of the learned coefficients in the different phases revealed that more learning leads to a higher spread in the coefficients over the subjects.
5.2. Field Trial
In the field trial part, the subjects used the experimental hearing instruments in their daily life for 6 weeks. They were requested to manipulate the instruments at will in order to maximize pleasantness of the listening experience. In Figure 11, we give an example of the (right ear) preference that is learned for subject 12. We visualize the learned coefficients by computing the noise reduction parameter that would result from steering by sounds with SNRs in the range of to 20 dB and power in the range of 50 to 90 dB. The color coding and the vertical axis of the learned surface correspond to the noise reduction parameter that would be predicted for a certain input sound. Because there is a nonlinear relation between computed SNR and power (in the features) and SNR and power of acoustic stimuli, the surface plot is slightly nonlinear as well. It can be seen that for high power and high SNR, a noise reduction of about 1 dB is obtained, which means that noise reduction is virtually inactive. For low power and low SNR, the noise reduction is almost equal to 7 dB, which means moderate noise reduction activity. The learned coefficients (and therefore also the noise reduction surfaces) show quite some variation among the subjects. Some are perfectly symmetric over the ears; others are quite asymmetric.
To assess this variation, we computed an estimate of the perceived "average noise reduction" over sounds ranging from SNR to 20 dB and power ranging from 50 to 90 dB. Sounds in this range will be particularly relevant to the hearing impaired since below SNR of dB virtually no intelligibility is left, and above 20 dB there is not much noise to suppress. Similarly, sounds with power below 50 dB will be almost inaudible to the hearing impaired. We call this estimate the "effective offset"—an estimate of the environment-independent part of the preferred noise reduction in the relevant acoustic range. The estimate was obtained by sampling the learned surface uniformly over the relevant acoustic range and computing the mean noise reduction parameter. This was done separately for each ear of each subject. The effective offset for left and right ears of all subjects is shown in the scatter plot of Figure 12. For example, subject 12 has an effective offset of approximately 4 dB in the right ear. This is visible in Figure 11 as a center of gravity of 4 dB.
From Figure 12, most subjects exhibit more or less symmetric noise reduction preference. However, subjects 8 and 10 (and to a lesser extent subjects 7 and 12) show a fair amount of asymmetry, and all these four subjects preferred learned settings over default noise reduction in lab trial 2. The need for personalization becomes clear from Figure 12 as well since the learned average parameter preferences cover almost the full range of the noise reduction parameter.
5.3. Lab Test 2
Subjects from group A listened to 5 minutes of acoustic stimuli using hearing instruments containing the noise reduction settings that were learned in the field trial. The sounds were a subset of the sounds in the first lab test which exhibited large transitions in SNR and SPL, but they are reflective of typical hearing conditions. The same sound file was played again with default noise reduction settings of 8 dB in all environments to compare sound quality and speech perception. Group B did the same in opposite order. Subjects did not know when default or learned settings were administered. The subjects were asked which of the two situations led to the most preferred sound experience. Two out of ten subjects did not have a preference, three had a small preference for the learned noise reduction settings, and five had a large preference for learned noise reduction settings (so 80% of the subjects had an overall preference for the learned settings). All subjects in the "majority group" in our trial judged the sound quality of the learned settings as "better" (e.g., "warmer sound" or "less effort to listen to it"), and seven out of eight felt that speech perception was better with learned settings. Nobody reported any artifacts of using the learning algorithm.
When looking more closely into the learned surfaces of all subjects, more than half of the subjects who preferred learned over default settings experienced a significantly sloping surface over the relevant acoustic range. The black dots on the surface of Figure 11 denote the sounds that have been used in the stimulus of the second lab test. From the position of these dots, we observe that during the second lab test, subject 12 experienced a noise reduction that changed considerably with the type of sound. We conjecture that the preference with respect to the default noise reduction setting is partly caused by the personalized environmental steering of the gain depth parameter.
By comparing the results of a final speech perception in noise task to those of the initial speech perception task in the initial fitting, it was concluded that the learned settings have no negative effect on conversational speech perception in noise. In fact, a lower speech reception threshold in noise was found with learned settings. However, a confounding factor is the prolonged use of new hearing instruments which may explain part of the improved intelligibility with learned settings.
In our approach to online personalization, an optional offline feature selection stage is included to enable more efficient learning during hearing aid use. From our feature selection experiments on synthetic data, we conclude that variational backfitting (VB) is a useful method for doing accurate regression and feature selection at the same time, provided that sample sizes are moderate to high and computation time is not an issue. Based on our preference data experiment, we selected the features of Power and WSNR for an experimental online learning algorithm. For one of the users, either the sample size was too low, his preference was too noisy, or the linearity assumption of the model might not hold. In our approach, we expect model mismatch (e.g., departure from linearity of the user's internal preference model) to show up as increased adjustment noise. Hence, a user who will never be fully satisfied with the linear mapping between features and noise reduction parameters because of model mismatch is expected to end up with a low learning rate (in the limit of many ongoing adjustments).
Our online learning algorithm can be looked upon as an interactive regression procedure. In the past, work on interactive curve fitting has been reported (e.g., see ). However, this work has limited value for hearing aid application since it requires an expensive library optimization procedure (like Nelder-Mead optimization) and probing of the user for ranking of parameter settings. In online settings, the user chooses the next listening experiment (the next parameter-feature setting for which a consent is given) rather than the learning algorithm. However, in the same spirit as this method, one may want to interpret a consent moment as a "ranking" of a certain parameter-feature setting at consent over a different setting at the preceding dissent moment. The challenge is then to absorb such rankings in an incremental, computationally efficient, and robust fashion. Indeed, we think that our approach to learning control can be adopted to other protocols (like learning from explicit dissent) and other user interfaces. Our aim is to embed the problem in a general framework for optimal Bayesian incremental fitting [14, 15], where a ranking of parameter values is used to incrementally train a user preference model.
In our second lab test, 80% of the subjects preferred learned over default settings. This is consistent with the findings by Zakis  who performed (semi-) online personalization of compressor gains using a standard least-squares method. Subjects had to confirm adjustments to a hearing aid as explicit training data, and after at least 50 "votes" an update to the gains was computed and applied. In two trials, subjects were asked to compare two settings of the aid during their daily life, where one setting was "some good initial setting" and the other was the "learned setting." The majority of the subjects preferred learned settings (70% of the subjects in the first trial, 80% in the second).
In recent work , Zakis et al. extended their personalization method to include noise suppression. Using the same semi-on-line learning protocol as before, a linear regression from sound pressure level and modulation depth to gain was performed. This was done for three different frequency (compression) bands separately by letting the control wheel operate in three different modes, in a cyclical manner. Modulation depth is used as an SNR estimate in each band, and by letting the gain in a band be steered with SNR, a trainable noise suppression can be obtained. Zakis et al. concluded that the provision of trained noise suppression did not have a significant additional effect on the preference for trained settings.
Although their work clearly demonstrates the potential of online hearing aid personalization, there are some issues that may prevent a successful practical application. First, their noise suppression personalization comes about by making per-band gains depend on per-band SNR. This requires a "looping mode implementation" of their learning control, where different bands are trained one after the other. This limits the amount of spectral resolution of the trainable noise suppression gain curve. In our approach, a 17-band gain curve is determined by a noise reduction method based on spectral subtraction, and we merely personalize an "aggressiveness" handle as a function of input power and weighted SNR. Apparently, a perceptual benefit may be obtained from such a learning noise control.
Furthermore, the explicit voting action and the looping mode of the gain control in  can make acceptance in the real world more difficult. We designed our learning control in such a way that it can be trained by using the hearing aid in the same way as a conventional hearing aid with control wheel. Further, in  environmental features have to be logged for at least 50 user actions, and additional updating requires a history of 50 to 256 votes, which limits the practicality of the method. Many users operate a control wheel for only a couple of times per day; so real-world learning with these settings may require considerable time before convergence is reached. In our approach, we learn incrementally from every user action, allowing fast convergence to preferred settings and low computational complexity. This is important for motivating subjects to operate the wheel for a brief period of time and then "set it and forget it" for the remainder of the usage. The faster reaction time of our algorithm comes at the expense of more uncertainty during each update, and by using a consistency tracker we avoid large updates when the user response contains a lot of uncertainty.
Interestingly, Zakis et al. found several large asymmetries between trained left and right steering coefficients, which they attribute to symmetric gain adjustments with highly asymmetric SPL estimates. We also found some asymmetric preferences in noise reduction. It is an open question whether these asymmetries are an artifact of the asymmetries in left and right sound fields or they reflect an actual preference for asymmetric settings with the user.
We described a new approach to online personalization of hearing instruments. Based on a linear mapping from acoustic features to user preferences, we investigated efficient feature selection methods and formulated the learning problem as the online maximization of the expected user utility. We then implemented an algorithm for online personalization on an experimental hearing aid, where we made use of the features that were selected in an earlier listening test. In a patient trial, we asked 10 hearing impaired subjects to use the experimental hearing aid in their daily life for six weeks. We then asked each patient to choose between the learned hearing aid settings and a (reasonable) default setting of the instrument. In this blind laboratory test, 80% of the subjects chose the learned settings, and nobody reported any artifacts of using the learning algorithm.
Launer S, Moore BCJ: Use of a loudness model for hearing aid fitting—V: on-line gain control in a digital hearing aid. International Journal of Audiology 2003,42(5):262-273. 10.3109/14992020309078345
Zakis JA: A trainable hearing aid, Ph.D. thesis. University of Melbourne, Melbourne, Australia; 2003.
Minka T: From hidden Markov models to linear dynamical systems. Department of Electrical Engineering and Computer Science, MIT, Cambridge, Mass, USA; 1999.
Jazwinski AH: Stochastic Processes and Filtering Theory. Academic Press, New York, NY, USA; 1970.
D'Souza AA: Towards tractable parameter-free statistical learning, Ph.D. thesis. University of Southern California, Los Angeles, Calif, USA; 2004.
Hastie TJ, Tibshirani RJ: Generalized Additive Models. Chapman & Hall/CRC, Boca Raton, Fla, USA; 1990.
Tipping ME: Bayesian inference: an introduction to principles and practice in machine learning. In Advanced Lectures on Machine Learning. Springer, New York, NY, USA; 2003:41-62.
Ypma A, Özer S, van der Werf E, de Vries B: Bayesian feature selection for hearing aid personalization. Proceedings of the 17th IEEE Workshop on Machine Learning for Signal Processing (MLSP '07), August 2007, Thessaloniki, Greece 425-430.
Guyon I, Elisseeff A: An introduction to variable and feature selection. The Journal of Machine Learning Research 2003, 3: 1157-1182.
Kates JM: Digital Hearing Aids. Plural Publishing, San Diego, Calif, USA; 2008.
Pavlovic CV: Band importance functions for audiological applications. Ear and Hearing 1994,15(1):100-104. 10.1097/00003446-199402000-00012
Plomp R, Mimpen AM: Improving the reliability of testing the speech reception threshold for sentences. International Journal of Audiology 1979,18(1):43-52. 10.3109/00206097909072618
Dennis JE, Woods DJ: Interactive graphics for curve-tailoring. In New Computing Environments: Microcomputers in Large-Scale Computing. SIAM, Philadelphia, Pa, USA; 1987:123-129.
Heskes T, de Vries B: Incremental utility elicitation for adaptive personalization. Proceedings of the 17th Belgium-Netherlands Conference on Artificial Intelligence (BNAIC '05), October 2005, Brussels, Belgium 127-134.
Dijkstra TMH, Ypma A, de Vries B, Leenen JRGM: The learning hearing aid: common-sense reasoning in hearing aid circuits. The Hearing Review 2007, 40-51.
Zakis JA, Dillon H, McDermott HJ: The design and evaluation of a hearing aid with trainable amplification parameters. Ear and Hearing 2007,28(6):812-830. 10.1097/AUD.0b013e3181576738
The authors would like to thank Tjeerd Dijkstra for preparation of the sound stimuli, and they are grateful to him, Almer van den Berg, Jos Leenen and Rob de Vries for useful discussions. They would also like to thank Judith Verberne for assistance with the patient trials. All collaborators are affiliated with GN ReSound Group.
A. Data Generation
For evaluation of the feature selection methods, we generated artificial regression data according to the following procedure.
(1)Choose total number of features and number of irrelevant features . The number of relevant features is .
(2)Generate samples from a normal distribution of dimension . Pad the input vector with zero dimensions.
(3)Regression coefficients were drawn from a normal distribution, and coefficients with value were clipped to . The first coefficients were put to zero.
(4)(Optional) Choose number of redundant features . The number of relevant features is now . Take the relevant features , rotate them with a random rotation matrix, and add them as redundant features by substituting features .
(5)Outputs were generated according to the model; Gaussian noise was added at an SNR of 10.
(6)An independent test set was generated in the same manner, but the output noise was zero in this case (i.e., an infinite output SNR).
(7)In all experiments, inputs and outputs were scaled to zero mean and unit variance after the data generation procedure. Unnormalized weights were found by inversely transforming the weights found by the algorithms. The noise variance parameters and were initialized to , thus assuming a total output noise variance that is 0.5 initially. We noticed that initializing the noise variances to large values led to slow convergence with large sample sizes. Initializing to alleviated this problem.
B. Efficient Precomputation
The standard least-squares error of a linear predictor, using weight vector and ignoring a constant term for the output variance, is calculated by
where is the autocorrelation matrix defined as
and is the cross-correlation vector defined as
Finding the optimal weights for , using standard least-squares fitting, requires a well-conditioned invertible matrix , which we ensure using a custom-designed regularization technique of adding a small fraction to the diagonal elements of the correlation matrix. Here, refers to the number of samples and refers to the number of selected features in the dataset. Since the regularized matrix is a nonsingular symmetrical positive definite matrix, we can use a Choleski factorization, providing an upper triangular matrix satisfying the relation , to efficiently compute the least-squares solution
Moreover, since intermediate solutions of actual weight values are often unnecessary because it suffices to have an error measure for a particular subset (with auto- and cross-correlations and obtained by selecting corresponding rows and columns of and , with being the corresponding Choleski factorization), we can directly insert (B.4) into (B.1) to efficiently obtain the error on the training set using
Obtaining a Choleski factorization from scratch, to test a selection of features, requires a computational complexity of , and the subsequent matrix division then only requires . The total effective complexity of the algorithm is .