A novel approach for estimating initial sound level for speech reception threshold test

2.1Air-conduction audiometry

Audiometry is the gold standard testing method for evaluating hearing availability by sound [7]. Two types of audiometry exist: air conduction (AC) and bone conduction. Pure tone and speech tone are involved in AC audiometry. Pure tone is used to determine the type and degree of hearing loss [8, 9], while speech tone is used to examine the applicability of hearing aids [5]. This section introduces detailed descriptions of pure tone and speech audiometry for the AC test.

2.1.1Pure tone audiometry

PTA aims to find the lowest hearing levels with sine wave sounds, termed pure tones [3]. Using different frequencies of pure tones, the hearing thresholds of patients corresponding to the frequencies, known as PTTs, are measured [6]. The PTA procedure for measuring PTTs is shown in Fig. 1.

Figure 1.

Procedure of measuring pure tone thresholds.

First, a patient was instructed to wear an AC headset in a soundproof booth and press a feedback button upon hearing a sound. After that, the pure tone sounds with d⁢Bs at xHz was played, where d⁢Bs and xHz represent the stimulus sound level and the frequency of pure tone sound, respectively. Usually, d⁢Bs ranged from 0 to 100 dB, and xHz was one of six frequencies: 250, 500, 1 K, 2 K, 4 K, and 8 KHz. After playing the pure tone sound of a certain frequency xHz at a certain d⁢Bs, the response was inspected to determine whether the feedback button was pressed. Depending upon the response, the sound level d⁢Bs was adjusted by decreasing 10 dB or increasing 5 dB until the audiologist determined the d⁢Bs as the PTT for the xHz frequency, represented by 𝑃𝑇𝑇𝑥𝐻𝑧.

As the result of the PTA for a patient, two data sets corresponding to the left and the right ears, each with six PTTs, were obtained as described in the Eq. (1).

(1)

𝑃𝑇𝐴d→={𝑃𝑇𝑇0.5⁢Kd,…,𝑃𝑇𝑇4⁢Kd,𝑃𝑇𝑇8⁢Kd}

where d represents either the left or the right ear. This data set was used to find the initial sound level for further processing, as described in the following section.

2.1.3Speech audiometry

In human communication, it is important not only to hear pure tone frequencies, but also to hear words. Given that hearing aids represent a prevalent treatment for hearing loss, evaluating their applicability is essential and accomplished through speech audiometry. Among various tests in SA, the SRT test holds particular significance because it measures the lowest hearing level with speech. The procedure of the SRT test is shown in Fig. 2.

Figure 2.

Procedure of speech reception threshold test.

The patient was instructed to do a “listen and repeat” which means involving the vocal repetition of heard words [6]. The audiologist initiated the presentation of a word with d⁢Bs determined as the initial sound level using a PTAv equation, and evaluated the accuracy of the patient in repeating the word. The count of correctly repeated words among N was counted during the test with a set of N words. If the number exceeded half of N, the sound level d⁢Bs was adjusted by decreasing 5 dB Conversely, if the number fell below half of N, the sound level d⁢Bs was adjusted by increasing 10 dB. This adjustment and “listen and repeat” test iterated with a different set of N words until precisely half of N was correctly repeated. The sound level used at that test cycle was set to 𝑆𝑅𝑇t, representing the SRT value for the patient, and the test was terminated [5].

As described in Fig. 2, the “listen and repeat” process was time-consuming, and patients, especially those required to discern small sound words, easily experienced fatigue. Minimizing the number of adjustments was imperative to simplify the entire hearing test and reduce patient fatigue. Note that, currently a PTAv [10] is used. This study proposes a machine learning approach to predict a better starting sound level 𝑆𝑅𝑇p using PTTs.

2.21D convolutional neural network

A convolutional neural network (CNN) is a neural network-based machine learning technique [11]. This algorithm integrates a 2-dimensional convolution layer, performing feature extraction through convolutional filtering operations, with a multi-layer feed-forward neural network. Although it is commonly used for image classification and segmentation due to its 2-dimensional operations [12], it is also adaptable to 1-dimensional signal data using a 1D convolution layer, referred to as a 1D CNN [13]. The multi-layer feed-forward neural network, also known as a multi-layer perceptron (MLP), is connected after the convolution layer to produce output values [14]. The activation function for neurons in both layers is typically non-linear, such as ReLU and Sigmoid [11]. Parameters for both layers are learned through the error back-propagation rule [15].

5.1CNN model architecture

In this study, the CNN model showing the state-of-the-art performance in many fields, was used [13, 14]. Alternatively, any other model capable of accurately predicting 𝑆𝑅𝑇t as the output, with 𝑃𝑇𝐴→ as the input vector, can be employed. Specifically, we used 1D CNN which is suitable for 1D signal input data involving 6 PTT values. The input layer obviously had six neurons corresponding to the number of PTTs. The output layer had single neuron with linear activation function to produce the output 𝑆𝑅𝑇p.

The architecture and hyper-parameters of the CNN model are summarized in Table 2. Note that, the architecture can significantly impact performance. However, detailed optimization of architecture and training parameters, such as drop-out, stride, batch normalization, and pooling [12], was not the primary focus because the primary goal was to introduce a novel approach employing machine learning techniques to reduce hearing test time by determining an optimal starting sound level. Further investigation into the ideal architecture for the model is recommended for enhanced performance.

The flatten layer transformed the outputs of 16 filters, each containing six values, into a flattened vector format of 96 values. This flattened vector served as the input to the fully connected layer. The output neuron had linear activation function, representing 𝑆𝑅𝑇p with values ranging from 0 dB to 100 dB, as detailed in Eq. (6). Note that the ReLU can also be applied as the activation function for the output neuron.

where i corresponds to neurons in the last hidden layer of the fully connected layer. Note that o denotes the neuron output, w corresponds to the connection weight, and b represents the bias.

5.2Training the CNN model

We randomly split the 10,481-sample data into two subsets for training and testing in a 50:50 ratio. The 1D CNN model was trained with the training subset to minimize the mean absolute error (MAE) defined by the Eq. (7):

where n denotes the number of data samples, 𝐸𝑟𝑟𝑜𝑟𝑆𝑅𝑇i corresponds to the prediction errors for the i-th sample, 𝑆𝑅𝑇ti represents the final SRT value determined by the test procedure depicted in Fig. 2, and 𝑆𝑅𝑇pi corresponds to the SRT value predicted by the output neuron of the proposed model for the i-th sample. MAE can be substituted with alternative error functions, such as mean squared error (MSE) related to the least square approach.

The proposed 1D CNN model was trained using a learning rate of 0.0001, using the ADAM algorithm as the optimization method, which is widely used in deep learning fields. The maximum training epoch was set at 10,000, and the MAE on the test subset was evaluated at each epoch to select the model with the best performance on the test subset. We again note that the primary focus was not on fine-tuning the architecture and training parameters.

The training and testing episode was iterated ten times to eliminate dependencies on data separation and parameter initialization, each time employing new data separation and initial parameters. Subsequently, we computed assessments for each episode, including the best, worst, and average outcomes. Further details are explained in the subsequent section.

6.1Result of model train

Figure 3.

Error curves for training data (a) and testing data (b).

The training progress is visualized in Fig. 3 through the 𝑀𝐴𝐸𝐶𝑁𝑁 curves for the training subset (left) and the test subset (right). The curves depict that the training process was effective across all ten trials, exhibiting a continuous decrease in both training and testing curves without signs of overfitting or memorization issues. The training curves decreased despite halting the training process at 10,000 epochs. This suggests that further training could yield even better results with lower MAEs.

𝑀𝐴𝐸𝐶𝑁𝑁 values for the ten experimental trials are reported in Table 3. Although not explicitly presented in this table, the ranking of the training results did not necessarily match the ranking of the test results.

6.2Prediction comparison

MAE values were computed for the conventional approaches 3FA, W3FA, and W4FA, employing the following equation to facilitate a comparison with conventional approaches elucidated in Section 2:

where 𝑥𝐹𝐴i. represents the PTAv computed by any 3FA, W3FA, or W4FA approaches for the i-th sample. For example, W4FAi(x=W4) was computed by the Eq. (4) in Section 2. Table 4 presents the 𝑀𝐴𝐸𝑥𝐹𝐴 values and the ratio between 𝑀𝐴𝐸𝐶𝑁𝑁 and 𝑀𝐴𝐸𝑥𝐹𝐴 for the testing set. Note that the 𝑀𝐴𝐸𝐶𝑁𝑁 value for the testing set, obtained by our method, was 2.96 dB on average, as indicated in Table 3.

As observed in Table 4, our machine learning approach reduced the SRT test time because of lower 𝑀𝐴𝐸𝐶𝑁𝑁 implied a smaller number of adjustments and repetitions to reach the 𝑆𝑅𝑇ti value. All 𝑀𝐴𝐸𝑥𝐹𝐴 values were larger than 𝑀𝐴𝐸𝐶𝑁𝑁, even for the worst result (3.01 dB). Although the ratio may not precisely exactly reflect the number of repetitions, we claimed that our method reduced the time for the SRT test by up to 37.92% compared to the 3FA approach.

6.3Discussion

In hearing tests, the most time-consuming aspect is the SRT test, involving the repetitive “listen and repeat” process with continuous adjustments of sound levels based on the responses of the patient. Prolonged test durations pose significant challenges because the monotonous nature of the “listen-and-repeat” process may lead to patient non-cooperative, resulting in unreliable test outcomes and inaccurate designs for hearing aids. Additionally, extended test times can inconvenience other patients who must wait for their turn. Therefore, our proposed approach is a promising solution to reduce the overall testing time by accurately predicting the SRT closer to its final value.

We did not focus on finding the optimal machine learning model and its parameters, while this study empirically demonstrates the successful application of machine learning techniques to reduce hearing test times. Hence, those implementing our approach for enhanced results should conduct further investigations to find a superior model and its associated parameters. Moreover, all experiments were performed using a dataset not specifically collected for our approach, limiting the comparison with conventional methods solely to MAE values. However, the number of repetitions for adjusting the sound level for the “listen and repeat” process should be compared.