Speech and Language Impairment Detection by Means of AI-Driven Audio-Based Techniques Luca Corvitto1 , Lorenzo Faiella1 , Christian Napoli1 , Adriano Puglisi1 and Samuele Russo2 1 Department of Computer, Control and Management Engineering, Sapienza University of Rome, Via Ariosto 25, Roma, 00185, Italy 2 Department of Psychology, Sapienza University of Rome, Via dei Marsi 78, Roma, 00185, Italy Abstract Speech and Language Impairments (SLI) affect a large and heterogeneous group of people. With our work, we propose a novel, easy, and immediate detection tool to help diagnose people who suffer from SLI using speech audio signals, along with a new dataset containing English speakers affected by SLI. In this work, we experiment with feature extraction methods such as Mel Spectrogram and wav2vec 2.0, as well as classification methods such as SVM, CNN, and linear neural networks. We also work on data audio augmentation trying to overcome the very common limitations imposed by data scarcity in the medical field. The overall results indicate that the wav2vec 2.0 feature extractor, paired with a linear classifier, provides the best performance with a reasonably high accuracy of over 96%. Keywords SLI, AI, audio, healthcare, speech, learning disease, feature extraction, data augmentation 1. Introduction sification systems. Automatic classification technologies are widely applied in voice assistants [8], chatbots [9], The rapid development of the use of Artificial Intelligence smart safety devices [10, 11], and in different real-world (AI) techniques in a broad range of scientific fields has environments [12, 13, 14]. helped solve real-life problems, in particular, the new ad- Our project aims to conciliate these two worlds and vancements revolutionized a wide variety of areas such as design a Deep Learning (DL) model that can detect, from Natural Language Processing (NLP) [1], computer vision a given audio input, if the speaker could be affected by [2, 3], robotics and many more. Due to the huge volume a speech and language impairment. Individuals with a of medical data being generated worldwide, there is a Speech and Language Impairment (SLI), generally, de- clear need for efficient use of this information to bene- spite normal hearing, normal nonverbal intelligence, ad- fit health sectors around the world [4, 5]. The medical equate social functioning, and no obvious signs of brain community has taken strong notice of the potential of injury represent a heterogeneous group of people with these new technologies in AI. Machine learning (ML) significant difficulty in learning languages [15]. One of thrives in areas where there are lots of data, therefore the defining characteristics of SLI is speech disfluency, ML is one of the essential and most effective tools in more specifically impaired acquisition of pattern-based analyzing highly complex medical data [6]. For example, components in language, such as morphology, syntax, analyzing medical data originating from disease diagno- and some aspects of phonology such as stuttering. This sis with the aid and benefits given by these tools could commonly used definition leads to early hypotheses re- be a lot more financially efficient. In healthcare, it is also garding the etiology of SLI that an impaired language- vital that diseases are detected early on during diagnosis specific learning mechanism underlies language develop- and prognosis. The success of these AI methods has also ment and disorders [16, 17, 18]. This disorder is deemed spread across other domains, including speech recogni- “primary” or “specific” when there is no clear explana- tion and the music recommendation task [7]. Due to the tion for these lags in language skills, a defining charac- relevance of such systems in our day-to-day lives, there teristic of primary language disorder is that its cause is is an increasing need for effective and efficient audio clas- unknown [19]. Language disorders are also linked to a heightened risk for psychiatric concerns, attentional ICYRIME 2024: 9th International Conference of Yearly Reports on Informatics, Mathematics, and Engineering. Catania, July 29-August difficulties, social-behavioral problems, and learning dis- 1, 2024 abilities [20, 21]. Many current trends in audio signal pro- $ corvitto.1835668@studenti.uniroma1.it (L. Corvitto); cessing rely on data-driven machine learning approaches faiella.1835950@studenti.uniroma1.it (L. Faiella); to achieve state-of-the-art results [22, 23, 24]. However, cnapoli@diag.uniroma1.it (C. Napoli); puglisi@diag.uniroma1.it the quantity and quality of available data influences heav- (A. Puglisi); samuele.russo@uniroma1.it (S. Russo) ily the achieved performance for a task. Depending on  0000-0002-3336-5853 (C. Napoli); 0009-0007-6307-7194 (A. Puglisi); 0000-0002-1846-9996 (S. Russo) the specific task, as for our case study, such data can © 2024 Copyright for this paper by its authors. Use permitted under Creative Commons License often be hard to obtain and costly to label particularly in Attribution 4.0 International (CC BY 4.0). CEUR Workshop Proceedings http://ceur-ws.org ISSN 1613-0073 CEUR Workshop Proceedings (CEUR-WS.org) 19 CEUR ceur-ws.org Workshop ISSN 1613-0073 Proceedings Luca Corvitto et al. CEUR Workshop Proceedings 19–31 the audio domain. As a consequence, researchers often reduction [33]. DA is key when dealing with problems have to deal with datasets of insufficient size or quality. regarding audio signals because the Convolutional Neu- Usually, diagnosis of this type of problem is carried out ral Network (CNN) is the most widely used model in with human experts, with special in-loco tests [25, 26] audio applications and when faced with small datasets, or with the aid of tools such as electroencephalogram CNN’s capacity for information retention becomes a flaw; (EEG) [27]. We want to design an easy and accessible the models memorize the training data and lose perfor- model that can detect if a person could be affected by mance on new data [34, 35]. In addition to increasing an SLI without having to go through complex and time- generalization capabilities, the augmentation of data also consuming procedures. In this manner, such a model allows the designed system to improve data significance, could also be implemented in robots from a human-robot regardless of the available data samples [36, 37]. These interaction (HRI) perspective, allowing the machine to strategies include methods on raw audio signals, as well detect people with SLI and change its behavior and form as applying other techniques on samples converted into of interaction accordingly. spectrograms or even more complex approaches such This study proposes an analysis of a novel, yet simple, as interpolation and nonlinear mixing on the spectrum. approach of using exclusively audio recordings for SLI We will now list and briefly explain the most used audio detection. Specifically, in Section 2 we start by exploring augmentation techniques. the current literature, and then we will talk about the Pitch Shifting. The tone of each audio signal in the problems faced in collecting our data and how we handled dataset is lowered or raised by a factor preserving its them in section 3. After that, in Section 4, we will go duration. through an analysis of the techniques and models used to Time Stretching. The audio sample is slowed down or perform the detection, the trials and results we obtained sped up by a ratio without altering the pitch drastically. from them in Section 5, and then we will discuss the Time Shifting. Time is shifted to the left or to the right limits of our approach in Section 6. We will finally draw by a random factor or by a predetermined amount. our conclusions in Section 7. Volume Adjustment. The volume of the audio file is altered, there is a change in loudness, or sometimes a dynamic range compression is applied. 2. Related Works Noise addition. Noise is introduced into the samples, other than a simple random Gaussian noise there are In the ever-evolving landscape of computer science and many types of noises such as white noise [38], babble artificial intelligence, the domains of audio data augmen- noise, static noise [39], factory noise, etc. tations and feature extraction are undergoing very rapid SpeedUp. The signal is resampled at a preset sampling changes and revolutions thanks to groundbreaking re- rate and later returned at the original sampling rate, re- search and advancements. In the following sections, we sulting in a speed change. will delve into the story and explore the state of the art Filtering. Several kinds of filters are applied to the of these fields. input audio. Most of the common filters are band-pass, band-stop, high-pass, high-shelf, low-pass, low-shelf, and 2.1. Audio Data Augmentation peaking filters. One of the most important challenges in developing an This topic is so important that researchers also devel- efficient and effective audio classification system is ac- oped and designed methods that generate entirely new cessing a large and well-annotated dataset. One of the samples, for example with the aid of a Generative Adver- main obstacles in developing sound classifications is a sarial Network (GAN) in [40] people created new variants lack of a sufficient quantity of labeled data. This is due of the audio samples that already existed in their dataset to the following main reasons: class imbalance, data pri- and then utilized an evolutionary algorithm to search vacy issues, time constraints involved in data collection, the input domain to select the best-generated samples, in high dependency on expertise for effective annotation, this way they were able to generate audio in a controlled etc. [28, 29, 30] Data Augmentation (DA) is defined as the manner that contributed to an improvement in classifica- creation of new data by adding deformations to increase tion performance of the original task. One very recent the variety of the data so that these deformations do not DA method proposed by Google is SpecAugment [41], change their semantic value. It is well known that DA can in this method, the two-dimensional spectrum diagram improve the algorithm’s performance, tackle the issue is treated as an image with time on the horizontal axis of overfitting [31, 32], and improve the generalization and frequency on the vertical axis. Encoder-decoder net- ability of Deep Neural Networks (DNN); this happens be- works are becoming very popular in fields different from cause DA averages over the orbits of the group that keeps NLP, this is because they can convert a high-dimensional the data distribution invariant, which leads to variance input into a lower-dimensional vector in latent space, researchers in [42] have experimented with a Long Short 20 Luca Corvitto et al. CEUR Workshop Proceedings 19–31 Term Memory (LSTM) based auto-encoder to produce trons (MLP) were very useful in person identification us- artificial data. ing speech and breath sounds [53], Hidden Markov Mod- els (HMM) [54], logistic regression and linear discrimi- 2.2. Audio Feature Extraction and Models nant analysis [55] and others. Some studies exploited the effectiveness of multiple simpler methods with ensemble It should be noted that data augmentation is not the only methods such as random forests [56, 51], XgBoost [57], way to reduce overfitting and improve the generalization and so on. Unfortunately, considering the complexity ability of DL models. Model structure optimization, trans- of sound and the need to sometimes train an extremely fer learning, and One-shot and Zero-shot learning are sensitive classifier that can identify different represen- also known strategies that deal with overfitting from dif- tations of sound features, traditional ML still suffers in ferent aspects. We will now focus on the most common these kinds of tasks from having less complex models. In processing flow of audio classification: preprocessing the this case, the choice of DL methods has been proven to be original audio data, feature extraction, and feeding the more efficient. DL methods differ from traditional ones features into the DL model. Audio signals have very high because they can extract meaningful features from data dimensionality, so thousands of floating point values are through the application of a hierarchical structure [58] required to represent a short audio signal, raising the CNNs were able to achieve significant and more accurate need for exploring dimensionality reduction and feature training results [59]. People tried to combine the best of extraction methods. The degree of how great or poor a these two worlds by implementing hybrid methods, for model performs is also determined by the choice of fea- example, researchers merged an SVM and a GRU-RNN tures used feature representation is crucial to improve the in [60]. performance of learning algorithms in the sound classifi- cation task. One of the first features that comes to mind when thinking of an audio signal is the spectrogram, its 3. Dataset characteristics have been widely used by previous re- searchers in different domains of sound classification, In the medical field, in particular, regarding specific prob- such as heartbeat sounds to detect heart diseases [43]. lems such as the one presented in this paper, data is not Another method used to extract features implements the always freely available or available at all. This is mostly Mel-Frequency Cepstrum (MFC), which is a representa- due to privacy concerns [61, 62]. Another important rea- tion of the short-term power spectrum of a sound, based son, which is also related in some ways to privacy [62], on a linear cosine transform of a log power spectrum lies in the overall low level of digitization of healthcare on a nonlinear mel scale of frequency, where the Mel- information [63]; in fact, according to Gopal G. et al. Frequency Cepstral Coefficients (MFCC) were successful [64], healthcare has the lowest level of digital innova- in representing sounds for the detection of respiratory tion compared to other industries, such as media, finance, diseases [44]. Some methods that also use the MFC are insurance, and retail, contributing to limited growth of the long-mel [45], mel filter bank energy [46], inverted labor productivity. In addition to this, it is also worth MFCC [47], and many more. Although mel spectrogram noting that not every dataset containing the desired med- and MFCC are commonly used, people also implement ical information is also in the desired format, in which bag of audio words [48], Discrete Gabor Transform (DGT) case the only remaining option is to create an entirely audio image representation [49], ZCR, entropy of energy, new dataset from scratch, that is what we did. spectral centroid, spectral spread, spectral entropy [50], and so on. 3.1. Data Collection Classification is a common task in ML and pattern The process of collecting audio data is a pivotal phase recognition. DL methods applied in these tasks, such as in this research. For our dataset, we aimed to collect a CNN models, often do not perform as well as more tradi- sufficient amount of pure, non-multimodal, audio data in tional ML methods such as random forest, Adaboost, etc., a waveform representation. Audio data can be stored in especially in small data [51]. On the other hand, typical various formats, each with its characteristics, trade-offs, ML algorithms, such as ensemble classifiers have been and use cases. Common audio formats include Wave- shown to learn features better and adapt more with im- form Audio File Format (WAV), MPEG-1 Audio Layer proved generalization abilities even in the case of small 3 (MP3), Free Lossless Audio Codec (FLAC), and more. and imbalanced datasets. Over the past years, differ- These formats differ in terms of compression, quality, and ent ML algorithms have been used for detecting sound compatibility. For this study, we opt for the WAV format events and medical sounds, and the achieved results were [65], which is an uncompressed audio file format, devel- of great significance. Classifiers, such as Support Vec- oped by IBM and Microsoft, that efficiently stores audio tor Machine (SVM), have shown to be very effective in data in a waveform representation without any loss of in- sound classification tasks [52], also MultiLayer Percep- 21 Luca Corvitto et al. CEUR Workshop Proceedings 19–31 Table 1 Dataset samples Train Test SLI Healthy SLI Healthy Non-augmented 1010 1010 124 125 Time-shifted 893 1010 104 125 Time-stretched 1010 1010 124 125 Figure 1: Audio sample from our dataset Pitch-shifted 2020 2020 248 250 Noise-addition 1010 1010 124 125 Total 5943 6060 724 750 Dataset 12003 1474 formation from them, keeping just human speech sounds (with or without background noise). After that, we an- alyzed the time windows by dividing each one of them into smaller ones containing the speech of one single Figure 2: Audio sample with noise from our dataset person each. Even if there exist different tools available to detect human speech, considering the scarcity of data we suffer, we decided to perform this step manually to formation. Thanks to its characteristics, which guarantee be sure that the quality of our dataset is not affected. the highest amount of information for an audio signal, Secondly we split the time windows that we obtained WAV is the audio format used as input by wav2vec 2.0 in 3-second clips. We chose this length as a trade-off be- [66], a state-of-the-art speech model developed by the tween a sufficient length, to capture fluency information Facebook AI Research group (FAIR) that is one of the and a brief duration. Our decision was also based on the models used in this work. standard approach used in the state-of-the-art working The data collection process began with the identifi- with wav2vec 2.0 in these kinds of tasks [68, 69]. Then cation of audio samples containing English speakers af- these clips were saved in two different subsets, creat- fected by Speech and Language Impairment (SLI) orig- ing the Train and the Test set, ensuring that the same inating from different conditions. This diverse dataset speakers do not overlap in both datasets. was intentionally curated to optimize the performance Finally the acquired data was augmented to increase of SLI detection. By including speakers with a range of its dimension. We applied the following audio augmen- impairments, the model is exposed to a broad spectrum tations techniques: Time shifting, Time stretching, Pitch of speech patterns and anomalies, thereby enhancing its shifting, and Noise addition, using Gaussian noise. To do ability to accurately detect SLI in real-world applications. so, we used the python library audiomentations [70]. For To source such data, we turned to YouTube, a vast and Time shifting we resampled the time windows shifting user-friendly repository of video and audio content. The the starting time further by 1.5 seconds; For Time stretch- videos found were then converted into audio files in WAV ing we slowed down the speed of the audios by a ratio format using an online converter. of 0.8; For the Pitch shifting we both lowered and raised We finally paired the collected data with a subset of the pitch tone by a value of 3, obtaining for each clip two the LibriSpeech dataset [67] containing healthy English additional ones; Finally for the Noise addition, we added speakers only. a 0.01𝑚 amplitude Gaussian noise. Audio waveforms before and after noise addition are shown in Fig. 1 and Fig. 2. All the augmentation techniques were applied on 3.2. Data Preprocessing the original audio; Time shifting was directly applied on To feed the waveform signals to the model, we needed the time windows, while the other ones on the initial 3 to ensure that they were appropriately prepared and pro- seconds clips. cessed. Effective data preprocessing is fundamental to The number of samples in the created dataset is shown enhancing the model’s performance, as it directly im- in Table 1, while in Table 2 we collect the audio data pacts the model’s ability to extract meaningful patterns augmentation techniques used and their respective pa- and insights from raw input data. This was performed rameters. in different steps. Firstly we identified different time windows from each audio file to cut out unnecessary in- 22 Luca Corvitto et al. CEUR Workshop Proceedings 19–31 Table 2 4.1. Log Mel Spectrogram Pamateres used for augmentation methods The way humans hear frequencies in sound is known as Augmentations Parameters pitch, it is a subjective impression of the frequency. They Time-shifting shift = +1.5 seconds do not perceive frequencies linearly, on the contrary, hu- Time-stretching ratio = 0.8 mans are more sensitive to differences between lower Pitch-shifting shift = ± 3 tones frequencies than higher ones. For example, the differ- Noise-addition amplitude = 0.01 meters ence between audios of frequency 100𝐻𝑧 and 200𝐻𝑧 is way bigger than 1000𝐻𝑧 and 1100𝐻𝑧, even though the absolute difference is the same amount. Humans per- ceive sounds on a logarithmic scale rather than a linear scale. The Mel Scale [72] was developed to take this into account by conducting experiments with a large number of listeners. It is a scale of pitches, such that each unit is judged by listeners to be equal in pitch distance from the next. The human perception of the amplitude of a sound is called loudness, similarly to frequency, also loud- ness is heard logarithmically rather than linearly. The Decibel scale is used to measure the loudness of a sound, for example, a sound with an amplitude of 20𝐷𝑏 is 10 times louder than one with an amplitude of 10𝐷𝑏. We can see that, to deal with sound realistically, we need to use a logarithmic scale via the Mel Scale and the Decibel Scale when dealing with Frequencies and Amplitudes in our data. Spectrograms are generated from sound signals using Fourier Transforms. A Fourier Transform (FT) [73] Figure 3: Log Mel Spectrogram of a sample from our dataset is a mathematical formula that allows us to decompose the signal into its constituent frequencies and displays the amplitude of each frequency present in the signal. 3.3. Data Management Spectrograms are generated from sound signals using FTs. In other words, an FT converts the signal from the The dataset contains audio files in the WAV format, its time domain into the frequency domain, and the result is data is affected not only by its advantages but also by its called a spectrum. A spectrogram consists in dividing the drawbacks. The complete dataset, which comprehends sound signal into smaller time segments, then applying both original and augmented data, was too large to be the FT to each segment, and finally, the combination of loaded in an online manner using the original files. To these segments in a single plot is called spectrogram. A overcome this problem we loaded the data in batches Mel Spectrogram makes two important changes relative and concatenated them in subsets that were saved in to a regular spectrogram that plots frequency vs time: it the .arrow format [71], a columnar memory format for uses the Mel scale instead of frequency on the y-axis and flat and hierarchical data, organized for efficient analytic uses the Decibel scale instead of amplitude to indicate operations. In this way, large data can be saved, loaded, color. In Fig. 3 we can see a normalized version of the Mel and processed avoiding memory usage problems. spectrogram of one of the audios present in the dataset. 4. Models and Techniques Used 4.2. Wav2vec 2.0 The best way to approach a problem is to know deeply Wav2vec 2.0 [66] is an exceptional tool that learns pow- every factor that influences it and how the key compo- erful representations from speech mimicking the human nents work, after that, one can tackle it and try to capture learning experience. People start, in fact, since the early its essence with the maximum capabilities. In the follow- stages of their lives comprehending language without la- ing subsections, we present a brief description of the beled data, i.e. kids learn from listening to adults around techniques we used and the models we implemented. them. It is also able to outperform state-of-the-art models while using 100 times less labeled data, thus demonstrat- ing the feasibility of training without huge amounts of labeled data which is very hard to achieve in a field deal- ing with a complex medium such as audio. 23 Luca Corvitto et al. CEUR Workshop Proceedings 19–31 4.3. Classification Methods Classification is the part that stands out the most in an entire model because it outputs the labels that are used to compute the evaluation metrics, even though it is the most noticeable part of a model, in our case they are just the final piece of the puzzle since most of the work is done in the previous steps of the pipeline; still, we want to pay some attention to the type of classifiers we used in our work. Support Vector Machine (SVM) [74] is one of the Figure 4: Wav2vec 2.0 pipeline first algorithms learned by every ML expert, it is sim- ple yet it can achieve excellent results, especially with small amounts of data where other ML algorithms tend to have some difficulties. The objective of the support The model can be visualized in Fig. 4 and next, we will vector machine algorithm is to find a hyperplane in an describe its components. N-dimensional space (𝑁 − the number of features) that Multi-layer convolutional feature encoder. It distinctly classifies the data points. To separate the two consists of several blocks containing a temporal con- classes of data points, many possible hyperplanes could volution followed by layer normalization and a GELU be chosen. SVM finds a plane that has the maximum mar- activation function. gin, i.e. the maximum distance between data points of Context network. It follows the Transformer ar- both classes. Maximizing the margin distance provides chitecture, differently from a normal Transformer that some reinforcement so that future data points can be uses fixed positional embeddings, a convolutional layer classified with more confidence. The biggest difficulty is used instead, and it acts as a relative positional embed- encountered when testing the SVM is that even with low ding. The output of the convolution followed by a GELU amounts of data the model had memory issues, since au- is added to the inputs and then a layer normalization is dio features are extremely large and with multiple classes, applied. while SVM excels with data that has fewer classes, thus Quantization module. It discretizes the output of making it hard to fully exploit SVM’s strengths. the feature encoder to a finite set of speech represen- One of the best and most efficient methods to generate tations via product quantization. Product quantization labels from an ML model is adding a linear layer at the amounts to choosing quantized representations from mul- end of the pipeline, that is what we did with our wav2vec tiple codebooks and concatenating them. The Gumbel 2.0 feature extractor, we have included a linear classifier softmax enables choosing discrete codebook entries in a 𝑓 (𝑥𝑖 , 𝑊, 𝑏) = 𝑊 · 𝑥𝑖 + 𝑏 and we trained its weights to fully differentiable way. output two types of labels, one for people affected by a The feature encoder 𝑓 : 𝑋 → 𝑍 takes as input the raw SLI and one for the others. waveform 𝑋 and outputs the latent speech representa- Resnet34 is a very famous residual neural network tions 𝑧1 , ..., 𝑍𝑡 for 𝑇 time steps, then they are fed to the that was pre-trained on ImageNet-1k and was released transformer 𝑔 : 𝑍 → 𝐶 that captures information from by Microsoft [75], thanks to residual learning and skip the entire sequence and outputs context representations. connections this type of model can be much deeper than The output of the feature encoder is also discretized to normal convolutional neural networks. We decided to 𝑞𝑡 with a quantization module to represent the targets fine-tune this model with the features extracted with the in the self-supervised objective. During the model’s pre- log mel spectrogram from our dataset. training a part of the latent speech representations that are generated from the feature encoder are masked, and then the model learns the representations of speech au- 5. Results dio by solving a contrastive task, which requires iden- tifying the true quantized latent speech representation In this section, we will describe the different architectures for a masked time step within a set of distractors. After that we tested in detail and then we will comment on the pre-training on unlabeled speech, the model is fine-tuned obtained results. on labeled data with a Connectionist Temporal Classifi- cation (CTC) loss. 5.1. Architectures Our first approach was to use the wav2vec 2.0 model, in particular the pre-trained wav2vec2-base model from HuggingFace [76], to perform Feature Extraction on the 24 Luca Corvitto et al. CEUR Workshop Proceedings 19–31 pre-processed non-augmented dataset and then use a Table 3 SVM, the Support Vector Classifier (SVC) model from Parameters used to compute the Spectrogram scikit learn [77], to perform the classification process tak- Log Mel Spectrum Parameters ing the extracted features in input. As it was explained in the previous section 4, wav2vec2.0 takes a raw waveform Sample rate 22050 signal as input, 3 seconds clips in WAV format in our Windows length 2048 case, then extracts audio features from them following Hop length 512 what it had learned in its previous training. The extracted N mels 128 features were then standardized using the StandardScaler from scikit learn, removing the mean and scaling them to Table 4 unit variance. The standardization of a dataset is a com- Architectures Accuracy mon requirement for many ML estimators: they might behave badly if the individual features do not more or less Models Accuracy look like standard normally distributed data (e.g. Gaus- LASSO (Full Model) [78] 0.84 sian with 0 mean and unit variance). Finally, we fitted 1NN CHI Strategy [79] 0.8832 the SVM using a linear kernel. LMT BL Strategy [79] 0.9269 Using the SVM model as a classifier was our first at- MLP BL Strategy [79] 0.9013 tempt to cope with the limited number of samples at our NB BL Strategy [79] 0.9269 CNN [80] 0.8421 disposal. Once the dataset was augmented we ceased to use the SVM due to its intrinsic limitations at work- Our Models Accuracy ing with large datasets; so we opted for a complete DL Wav2vec2.0 + SVM 0.6627 approach. Wav2vec2.0 + FC 0.9661 For our second architecture, we substituted the classi- Log Mel Spectrogram + CNN 0.9362 fier head with a simple Fully Connected (FC), or linear, layer, keeping the wav2vec 2.0 model to perform the Feature Extraction, this time, on the augmented dataset. probably due to the magnitude of the feature space ex- We trained the model for 5 epochs through the Trainer tracted by the wav2vec 2.0 model. class by HuggingFace on a batch of 32 samples each, set- Using, instead, an augmented dataset together within ting the learning rate to 2𝑒 − 5 after a warm-up period a DL approach we manage to reach a very high value at a ratio of 0.1 and decreasing its value linearly till the of accuracy, the highest of our models. The wav2vec end of the training. 2.0 feature extractor, having enough data to work with, The last architecture tested was a CNN, more precisely managed to extract the key features and information resnet34, that received as input the log mel spectrogram needed to correctly identify which voice belongs to a of the audios and generated as output the labels of the healthy speaker or an impaired one. given audio. All the procedures to extract the spectro- The CNN model that was fine-tuned with Log Mel gram were carried on with the librosa library, firstly the Spectrum features achieved great accuracy in labeling sample was resampled with a new rate of 22050, then samples, unfortunately, through a more accurate analysis the mel spectrogram extracted was normalized and fi- of the confusion matrices shown in 5, 6, and 7, we dis- nally scaled. Regarding the CNN, only the last layer was covered that the number of false negatives is extremely modified, it was replaced with a linear layer that had two high compared to the false positives. In the medical field, output channels and the whole model was fine-tuned especially for tools helping with diagnosis, it is crucial without freezing the previous layers. Training was car- to have the smallest number of false negatives, since an ried out for 50 epochs, the learning rate started at 2𝑒 − 4 undetected disease is much worse than a false positive, and decayed by a factor of 10 every 10 epochs; the loss medical operators could be missing a lot of vital anoma- function used was the CrossEntropyLoss. All parameters lies and in time they will lose trust in the system. In used to compute the spectrum are shown in Table 3. our case recall is way more important than the preci- sion score, from Table 5 we can see that the CNN model 5.2. Evaluations reaches only a recall score of 0.85, on the other hand wav2vec 2.0 achieves a better recall and F1 Score. In Table 4 we show the accuracy of our architectures, compared with others architectures [78] As we can see, the first model is the one with the lowest score. This 6. Limitations and Future Works means that, despite the ability of the SVM to avoid over- fitting on the poor quantity of data provided, it cannot It is of critical importance to examine our achievements accurately detect the speakers affected by SLI. This is and acknowledge the constraints that affect our work. 25 Luca Corvitto et al. CEUR Workshop Proceedings 19–31 Table 5 Architectures overall performances +HDOWK\   Model F1 Score Precision Recall Wav2vec2.0 + SVM 0.6316 0.6923 0.5806 Wav2vec2.0 + FC 0.9655 0.9641 0.9668 Spectrogram + CNN 0.9187 0.9983 0.8508   6/, +HDOWK\ 6/, While our research has given promising results, the fol- lowing section delves into the limitations that shape our Figure 5: Wav2vec 2.0 + SVM confusion matrix results and sets a base for future possible improvements. 6.1. Limitations 6.2. Future Works The lack of data quantity and quality is one of our major constraints. The problem of data scarcity has already Future works should focus on the creation of a new been addressed in section 3 so we will now talk about dataset comprising people speaking different languages, quality. since it is not yet known, to our knowledge, whether flu- In the realm of ML and DL, it has been well docu- ency problems can be generalized in all languages and a mented that the issue of low-quality data and disparities wide age range, knowing that the features and the overall in data collection methodologies exacerbate the inherent characteristics of the voice between children and adults biases within the data when utilized for training algo- change in general, due to their anatomical differences rithms, a clear example is given by the societal or political [86]. biases reflected in word embeddings or large language Given the technological advancement in the field of models [81, 82]. This concern arises when the data col- generative audio with astonishing tools such as the au- lected for training purposes exhibits significant varia- dio manipulation software produced by ElevenLabs [87], tions in quality and collection techniques, resulting in which can clone voices, generate new ones, translate a heightened vulnerability to intrinsic biases within the them into other languages, and make them read texts, data. Such biases can subsequently propagate through new kinds of audio enhancement can be experimented the training process, influencing the performance and with, and although they cannot be used now, because fairness of ML and DL algorithms leading to further dis- they cannot replicate stuttering or other kinds of fluency parities and discrimination in the real world, due to the features that characterize people affected with SLI yet, accessibility to such tools [83, 84]. Particularly, in our they are promising tools to take into consideration for work, the collection of English speakers affected by SLI the near future. presents the limitation of containing mostly speakers with American accents. In real-world applications this 7. Conclusions can have negative effects on the model performance, for example, the algorithm could achieve higher and better This work proposes a novel approach to Speech and Lan- results with American people rather than with Mexican guage Impairment (SLI) detection, based solely on audio ones, or other English-speaking minority ethnic groups and AI audio-based techniques, together within an en- of people whose accent differs from the standard Ameri- tirely new dataset composed of English speakers affected can one [84]. by SLI. The results show that, even with some limitations Another limitation of our dataset is that it does not related to the scarcity of data available, Deep Learning contain children speakers. This is because finding such methods can achieve accurate estimations on healthy materials on the web is often difficult, and it is more or impaired speakers. In particular, wav2vec 2.0, with a difficult to create them from scratch due to the small Fully Connected layer as the classification head, reaches number of certified children affected by SLI and, since an accuracy of over 96% on our test set. Our findings also they are minors, due to more strict privacy concerns. confirm that data audio augmentation techniques are fun- The most used dataset in this field [85] consists of one damental to training Deep Learning models adequately. second clips of Czech speaking children, both healthy or affected by SLI. Although this dataset could be useful for the detection of SLI, it is limited to the Czech language and children speakers. This kind of limitation is common in the healthcare field, especially in SLI detection. 26 Luca Corvitto et al. CEUR Workshop Proceedings 19–31 doi:10.1161/CIRCRESAHA.115.306013, cited by: 44; All Open Access, Bronze Open Access, Green Open Access. +HDOWK\   [5] E. Iacobelli, V. Ponzi, S. Russo, C. Napoli, Eye- tracking system with low-end hardware: Devel- opment and evaluation, Information (Switzerland)   14 (2023). doi:10.3390/info14120644. 6/, [6] M. Woźniak, D. Połap, R. K. Nowicki, C. Napoli, +HDOWK\ 6/, G. Pappalardo, E. Tramontana, Novel approach toward medical signals classifier, in: Proceedings of the International Joint Conference on Neural Figure 6: Wav2vec 2.0 + FC confusion matrix Networks, volume September 2015, 2015. doi:10. 1109/IJCNN.2015.7280556. [7] P. Zinemanas, M. Rocamora, M. Miron, F. Font, X. Serra, An interpretable deep learning model for automatic sound classification, Electronics 10 +HDOWK\   (2021). URL: https://www.mdpi.com/2079-9292/10/ 7/850. doi:10.3390/electronics10070850. [8] M. Azimi, U. Roedig, Room Identifica-   tion with Personal Voice Assistants (Ex- 6/, tended Abstract), 2022, pp. 317–327. doi:10.1007/978-3-030-95484-0_19. +HDOWK\ 6/, [9] J. Kapočiūṫe-Dzikieṅe, A domain-specific generative chatbot trained from little Figure 7: Log Mel Spectrogram + CNN confusion matrix data, Applied Sciences 10 (2020). URL: https://www.mdpi.com/2076-3417/10/7/2221. doi:10.3390/app10072221. References [10] S. Shah, Z. Tariq, Y. Lee, Audio iot analytics for home automation safety, 2018, pp. 5181–5186. [1] A. Le Glaz, Y. Haralambous, D.-H. Kim-Dufor, doi:10.1109/BigData.2018.8622587. P. Lenca, R. Billot, T. C. Ryan, J. Marsh, J. DeVylder, [11] F. Fiani, S. Russo, C. Napoli, An advanced solu- M. Walter, S. Berrouiguet, C. Lemey, Machine tion based on machine learning for remote emdr learning and natural language processing in mental therapy, Technologies 11 (2023). doi:10.3390/ health: Systematic review, J Med Internet Res 23 technologies11060172. (2021) e15708. [12] S. Gholizadeh, Z. Leman, B. T. Baharudin, A review [2] D. I. Patrício, R. Rieder, Computer vision and of the application of acoustic emission technique in artificial intelligence in precision agriculture for engineering, Structural Engineering and Mechan- grain crops: A systematic review, Comput- ics 54 (2015) 1075–1095. doi:10.12989/sem.2015. ers and Electronics in Agriculture 153 (2018) 69– 54.6.1075. 81. URL: https://www.sciencedirect.com/science/ [13] H. Lozano, I. Hernáez, A. Picón, J. Camarena, article/pii/S0168169918305829. doi:https://doi. E. Navas, Audio classification techniques in home org/10.1016/j.compag.2018.08.001. environments for elderly/dependant people, in: [3] I. E. Tibermacine, A. Tibermacine, W. Guettala, K. Miesenberger, J. Klaus, W. Zagler, A. Karsh- C. Napoli, S. Russo, Enhancing sentiment anal- mer (Eds.), Computers Helping People with Special ysis on seed-iv dataset with vision transformers: Needs, Springer Berlin Heidelberg, Berlin, Heidel- A comparative study, in: ACM International berg, 2010, pp. 320–323. Conference Proceeding Series, 2023, p. 238 – 246. [14] D. T. Blumstein, D. J. Mennill, P. Clemins, L. Girod, doi:10.1145/3638985.3639024. K. Yao, G. Patricelli, J. L. Deppe, A. H. Krakauer, [4] S. B. Scruggs, K. Watson, A. I. Su, H. Hermjakob, C. Clark, K. A. Cortopassi, S. F. Hanser, B. Mc- J. R. Yates, M. L. Lindsey, P. Ping, Harnessing Cowan, A. M. Ali, A. N. G. Kirschel, Acoustic the heart of big data, Circulation Research 116 monitoring in terrestrial environments using (2015) 1115 – 1119. URL: https://www.scopus.com/ microphone arrays: applications, technologi- inward/record.uri?eid=2-s2.0-84930702552&doi= cal considerations and prospectus, Journal of 10.1161%2fCIRCRESAHA.115.306013&partnerID= Applied Ecology 48 (2011) 758–767. URL: https: 40&md5=4244ed52d2f51fa5c08f02ca67e4103e. //besjournals.onlinelibrary.wiley.com/doi/abs/10. 27 Luca Corvitto et al. CEUR Workshop Proceedings 19–31 1111/j.1365-2664.2011.01993.x. doi:https://doi. Disabil. 52 (2019) 351–365. org/10.1111/j.1365-2664.2011.01993.x. [26] J. C. Lee, J. B. Tomblin, Reinforcement learning in [15] D. V. M. Bishop, Uncommon understanding: Devel- young adults with developmental language impair- opment and disorders of language comprehension ment, Brain Lang. 123 (2012) 154–163. in children, Psychology Press/Erlbaum (UK) Taylor [27] R. A. Ahire, Nitin, A. Wagh, Eeg based identifica- and Francis, 1997. tion of learning disabilities using machine learning [16] H. CLAHSEN, The grammatical characterization of algorithms, J Neurol Disord (2022). developmental dysphasia 27 (1989) 897–920. URL: [28] O. O. Abayomi-Alli, R. Damaševičius, A. Qazi, https://doi.org/10.1515/ling.1989.27.5.897. doi:doi: M. Adedoyin-Olowe, S. Misra, Data augmenta- 10.1515/ling.1989.27.5.897. tion and deep learning methods in sound classi- [17] M. L. Rice, K. Wexler, P. L. Cleave, Specific language fication: A systematic review, Electronics 11 (2022). impairment as a period of extended optional infini- URL: https://www.mdpi.com/2079-9292/11/22/3795. tive, Journal of Speech, Language, and Hearing doi:10.3390/electronics11223795. Research 38 (1995) 850–863. URL: https://pubs.asha. [29] C. Napoli, G. Pappalardo, E. Tramontana, Using org/doi/abs/10.1044/jshr.3804.850. doi:10.1044/ modularity metrics to assist move method refactor- jshr.3804.850. ing of large systems, in: Proceedings - 2013 7th [18] H. K. van der Lely, Domain-specific cognitive International Conference on Complex, Intelligent, systems: insight from grammatical-sli, Trends and Software Intensive Systems, CISIS 2013, 2013, in Cognitive Sciences 9 (2005) 53–59. URL: https: p. 529 – 534. doi:10.1109/CISIS.2013.96. //doi.org/10.1016/j.tics.2004.12.002. doi:10.1016/ [30] D. Połap, M. Woźniak, C. Napoli, E. Tramontana, j.tics.2004.12.002. Real-time cloud-based game management system [19] D. V. M. Bishop, Ten questions about terminology via cuckoo search algorithm, International Journal for children with unexplained language problems, of Electronics and Telecommunications 61 (2015) Int. J. Lang. Commun. Disord. 49 (2014) 381–415. 333 – 338. doi:10.1515/eletel-2015-0043. [20] J. H. Beitchman, B. Wilson, E. B. Brownlie, H. Wal- [31] Z. Mushtaq, S.-F. Su, Q.-V. Tran, Spectral im- ters, A. Inglis, W. Lancee, Long-term consistency in ages based environmental sound classification speech/language profiles: II. behavioral, emotional, using cnn with meaningful data augmentation, and social outcomes, J. Am. Acad. Child Adolesc. Applied Acoustics 172 (2021) 107581. URL: Psychiatry 35 (1996) 815–825. https://www.sciencedirect.com/science/article/pii/ [21] T. L. Stanton-Chapman, L. M. Justice, L. E. Skibbe, S0003682X2030685X. doi:https://doi.org/10. S. L. Grant, Social and behavioral charac- 1016/j.apacoust.2020.107581. teristics of preschoolers with specific language [32] M. Woźniak, D. Połap, C. Napoli, E. Tramontana, impairment, Topics in Early Childhood Spe- Graphic object feature extraction system based on cial Education 27 (2007) 98–109. URL: https://doi. cuckoo search algorithm, Expert Systems with Ap- org/10.1177/02711214070270020501. doi:10.1177/ plications 66 (2016) 20 – 31. doi:10.1016/j.eswa. 02711214070270020501. 2016.08.068. [22] J. Wagner, D. Schiller, A. Seiderer, E. André, Deep [33] S. Chen, E. Dobriban, J. Lee, Invariance reduces learning in paralinguistic recognition tasks: Are variance: Understanding data augmentation in deep hand-crafted features still relevant?, in: Inter- learning and beyond, ArXiv abs/1907.10905 (2019). speech, 2018. URL: https://api.semanticscholar.org/ URL: https://api.semanticscholar.org/CorpusID: CorpusID:52192644. 198895147. [23] Y. Tokozume, T. Harada, Learning environmental [34] C. Shorten, T. M. Khoshgoftaar, A survey sounds with end-to-end convolutional neural net- on image data augmentation for deep learning, work, in: 2017 IEEE International Conference on Journal of Big Data 6 (2019) 60. URL: https:// Acoustics, Speech and Signal Processing (ICASSP), doi.org/10.1186/s40537-019-0197-0. doi:10.1186/ 2017, pp. 2721–2725. doi:10.1109/ICASSP.2017. s40537-019-0197-0. 7952651. [35] C. Napoli, G. Pappalardo, E. Tramontana, Z. Marsza- [24] J. Lee, J. Park, K. L. Kim, J. Nam, Samplecnn: End- lek, D. Polap, M. Wozniak, Simplified firefly algo- to-end deep convolutional neural networks using rithm for 2d image key-points search, in: IEEE SSCI very small filters for music classification, Applied 2014 - 2014 IEEE Symposium Series on Computa- Sciences 8 (2018). URL: https://www.mdpi.com/ tional Intelligence - CIHLI 2014: 2014 IEEE Sym- 2076-3417/8/1/150. doi:10.3390/app8010150. posium on Computational Intelligence for Human- [25] L. M. Justice, W.-Y. Ahn, J. A. R. Logan, Identifying Like Intelligence, Proceedings, 2014. doi:10.1109/ children with clinical language disorder: An appli- CIHLI.2014.7013395. cation of machine-learning classification, J. Learn. [36] A. Greco, N. Petkov, A. Saggese, M. Vento, Aren: 28 Luca Corvitto et al. CEUR Workshop Proceedings 19–31 A deep learning approach for sound event recog- [45] Y. Leng, W. Zhao, C. Lin, C. Sun, R. Wang, Q. Yuan, nition using a brain inspired representation, IEEE D. Li, Lda-based data augmentation algorithm for Transactions on Information Forensics and Security acoustic scene classification, Knowledge-Based Sys- 15 (2020) 3610–3624. doi:10.1109/TIFS.2020. tems 195 (2020) 105600. doi:10.1016/j.knosys. 2994740. 2020.105600. [37] D. Połap, M. Woźniak, C. Napoli, E. Tramontana, [46] V. M. Praseetha, P. P. Joby, Speech emotion recog- R. Damaševičius, Is the colony of ants able to rec- nition using data augmentation, International ognize graphic objects?, Communications in Com- Journal of Speech Technology 25 (2022) 783–792. puter and Information Science 538 (2015) 376 – 387. URL: https://doi.org/10.1007/s10772-021-09883-3. doi:10.1007/978-3-319-24770-0_33. doi:10.1007/s10772-021-09883-3. [38] Z. Mushtaq, S.-F. Su, Environmental sound [47] S. Lalitha, D. Gupta, M. Zakariah, Y. A. Alotaibi, classification using a regularized deep convolu- Investigation of multilingual and mixed-lingual tional neural network with data augmentation, emotion recognition using enhanced cues with Applied Acoustics 167 (2020) 107389. URL: data augmentation, Applied Acoustics 170 https://www.sciencedirect.com/science/article/pii/ (2020) 107519. URL: https://www.sciencedirect. S0003682X2030493X. doi:https://doi.org/10. com/science/article/pii/S0003682X2030623X. 1016/j.apacoust.2020.107389. doi:https://doi.org/10.1016/j.apacoust. [39] O. Novotny, O. Plchot, O. Glembek, J. H. Cer- 2020.107519. nocky, L. Burget, Analysis of dnn speech sig- [48] M. Schmitt, C. Janott, V. Pandit, K. Qian, C. Heiser, nal enhancement for robust speaker recognition, W. Hemmert, B. Schuller, A bag-of-audio-words Computer Speech and Language 58 (2019) 403– approach for snore sounds’ excitation localisation, 421. URL: https://www.sciencedirect.com/science/ in: Speech Communication; 12. ITG Symposium, article/pii/S0885230818303607. doi:https://doi. 2016, pp. 1–5. org/10.1016/j.csl.2019.06.004. [49] H. Lachambre, B. Ricaud, G. Stempfel, B. Torrésani, [40] S. Mertes, A. Baird, D. Schiller, B. W. Schuller, C. Wiesmeyr, D. Onchis-Moaca, Optimal window E. André, An evolutionary-based generative ap- and lattice in gabor transform. application to audio proach for audio data augmentation, in: 2020 IEEE analysis, in: 2015 17th International Symposium 22nd International Workshop on Multimedia Signal on Symbolic and Numeric Algorithms for Scientific Processing (MMSP), 2020, pp. 1–6. doi:10.1109/ Computing (SYNASC), 2015, pp. 109–112. doi:10. MMSP48831.2020.9287156. 1109/SYNASC.2015.25. [41] D. S. Park, W. Chan, Y. Zhang, C.-C. Chiu, [50] E. Garcia-Ceja, M. Riegler, A. K. Kvernberg, J. Tor- B. Zoph, E. D. Cubuk, Q. V. Le, SpecAug- resen, User-adaptive models for activity and ment: A simple data augmentation method emotion recognition using deep transfer learn- for automatic speech recognition, in: Inter- ing and data augmentation, User Modeling and speech 2019, ISCA, 2019. URL: https://doi.org/10. User-Adapted Interaction 30 (2020) 365–393. URL: 21437%2Finterspeech.2019-2680. doi:10.21437/ https://doi.org/10.1007/s11257-019-09248-1. doi:10. interspeech.2019-2680. 1007/s11257-019-09248-1. [42] E. K. Wang, J. Yu, C.-M. Chen, S. Kumari, J. J. [51] H. Ykhlef, F. Ykhlef, S. Chiboub, Experimental P. C. Rodrigues, Data augmentation for in- design and analysis of sound event detection sys- ternet of things dialog system, Mobile Net- tems: Case studies, in: 2019 6th International Con- works and Applications 27 (2022) 158–171. URL: ference on Image and Signal Processing and their https://doi.org/10.1007/s11036-020-01638-9. doi:10. Applications (ISPA), 2019, pp. 1–6. doi:10.1109/ 1007/s11036-020-01638-9. ISPA48434.2019.8966798. [43] T. Koike, K. Qian, B. W. Schuller, Y. Yamamoto, [52] S. Lalitha, D. Gupta, M. Zakariah, Y. A. Alotaibi, Transferring cross-corpus knowledge: An investi- Investigation of multilingual and mixed-lingual gation on data augmentation for heart sound classi- emotion recognition using enhanced cues with fication, in: 2021 43rd Annual International Confer- data augmentation, Applied Acoustics 170 ence of the IEEE Engineering in Medicine & Biology (2020) 107519. URL: https://www.sciencedirect. Society (EMBC), IEEE, 2021. com/science/article/pii/S0003682X2030623X. [44] V. Basu, S. Rana, Respiratory diseases recognition doi:https://doi.org/10.1016/j.apacoust. through respiratory sound with the help of deep 2020.107519. neural network, in: 2020 4th International Confer- [53] V.-T. Tran, W.-H. Tsai, Stethoscope-sensed speech ence on Computational Intelligence and Networks and breath-sounds for person identification with (CINE), 2020, pp. 1–6. doi:10.1109/CINE48825. sparse training data, IEEE Sensors Journal 20 (2020) 2020.234388. 848–859. doi:10.1109/JSEN.2019.2945364. 29 Luca Corvitto et al. CEUR Workshop Proceedings 19–31 [54] T. A. M. Celin, T. Nagarajan, P. Vijayalakshmi, [63] T. M. Stoumpos AI, Kitsios F, Digital transforma- Data augmentation using virtual microphone array tion in healthcare: Technology acceptance and its synthesis and multi-resolution feature extraction applications, Int J Environ Res Public Health (2023). for isolated word dysarthric speech recognition, doi:10.3390/ijerph20043407. IEEE Journal of Selected Topics in Signal Process- [64] G. Gopal, C. Suter-Crazzolara, L. Toldo, W. Eber- ing 14 (2020) 346–354. doi:10.1109/JSTSP.2020. hardt, Digital transformation in healthcare – 2972161. architectures of present and future information [55] J. Ye, T. Kobayashi, M. Murakawa, Urban sound technologies, Clinical Chemistry and Labora- event classification based on local and global tory Medicine (CCLM) 57 (2019) 328–335. URL: features aggregation, Applied Acoustics 117 https://doi.org/10.1515/cclm-2018-0658. doi:doi: (2017) 246–256. URL: https://www.sciencedirect. 10.1515/cclm-2018-0658. com/science/article/pii/S0003682X16302274. [65] Wave file format specification, ???? URL: doi:https://doi.org/10.1016/j.apacoust. https://www.mmsp.ece.mcgill.ca/Documents/ 2016.08.002, acoustics in Smart Cities. AudioFormats/WAVE/WAVE.html. [56] V. Ramesh, K. Vatanparvar, E. Nemati, V. Nathan, [66] A. Baevski, H. Zhou, A. Mohamed, M. Auli, M. M. Rahman, J. Kuang, CoughGAN: Generating Wav2vec 2.0: A framework for self-supervised synthetic coughs that improve respiratory disease learning of speech representations, in: Proceed- classification(), Annu Int Conf IEEE Eng Med Biol ings of the 34th International Conference on Neural Soc 2020 (2020) 5682–5688. Information Processing Systems, NIPS’20, Curran [57] N. Yella, B. Rajan, Data augmentation using gan for Associates Inc., Red Hook, NY, USA, 2020. sound based covid 19 diagnosis, in: 2021 11th IEEE [67] V. Panayotov, G. Chen, D. Povey, S. Khudanpur, International Conference on Intelligent Data Acqui- Librispeech: An asr corpus based on public do- sition and Advanced Computing Systems: Technol- main audio books, in: 2015 IEEE International Con- ogy and Applications (IDAACS), volume 2, 2021, ference on Acoustics, Speech and Signal Process- pp. 606–609. doi:10.1109/IDAACS53288.2021. ing (ICASSP), 2015, pp. 5206–5210. doi:10.1109/ 9660990. ICASSP.2015.7178964. [58] H. Lee, J. Lee, Neural network prediction of sound [68] T. Grósz, D. Porjazovski, Y. Getman, S. Kadiri, M. Ku- quality via domain knowledge-based data augmen- rimo, Wav2vec2-based paralinguistic systems to tation and bayesian approach with small data sets, recognise vocalised emotions and stuttering, in: Mechanical Systems and Signal Processing 157 Proceedings of the 30th ACM International Con- (2021) 107713. URL: https://www.sciencedirect.com/ ference on Multimedia, MM ’22, Association for science/article/pii/S0888327021001084. doi:https: Computing Machinery, New York, NY, USA, 2022, //doi.org/10.1016/j.ymssp.2021.107713. p. 7026–7029. URL: https://doi.org/10.1145/3503161. [59] D. Koszewski, B. Kostek, Musical instrument tag- 3551572. doi:10.1145/3503161.3551572. ging using data augmentation and effective noisy [69] J. Liu, A. Wumaier, D. Wei, S. Guo, Automatic data processing, Journal of the Audio Engineer- speech disfluency detection using wav2vec2. 0 for ing Society 68 (2020) 57–65. doi:10.17743/jaes. different languages with variable lengths, Applied 2019.0050. Sciences 13 (2023) 7579. [60] Z. Zhang, J. Han, K. Qian, C. Janott, Y. Guo, [70] I. Jordal, A. Tamazian, E. T. Chourdakis, An- B. Schuller, Snore-GANs: Improving automatic gonin, askskro, N. Karpov, T. Dhyani, O. Sar- snore sound classification with synthesized data, ioglu, kvilouras, E. Berk, F. Mirus, J.-Y. Lee, IEEE J Biomed Health Inform 24 (2019) 300–310. K. Choi, MarvinLvn, SolomidHero, T. Alum, [61] K. P. Seastedt, P. Schwab, Z. O’Brien, E. Wakida, iver56/audiomentations: v0.33.0 (????). doi:10. K. Herrera, P. G. F. Marcelo, L. Agha-Mir-Salim, X. B. 5281/zenodo.7010042. Frigola, E. B. Ndulue, A. Marcelo, L. A. Celi, Global [71] N. Richardson, I. Cook, N. Crane, D. Dun- healthcare fairness: We should be sharing more, nington, R. François, J. Keane, D. Moldovan- not less, data, PLOS Digital Health 1 (2022) 1–13. Grünfeld, J. Ooms, Apache Arrow, ar- URL: https://doi.org/10.1371/journal.pdig.0000102. row: Integration to ’Apache’ ’Arrow’, doi:10.1371/journal.pdig.0000102. 2023. Https://github.com/apache/arrow/, [62] M. Paul, L. Maglaras, M. A. Ferrag, I. Almomani, https://arrow.apache.org/docs/r/. Digitization of healthcare sector: A study on [72] B. Truax, Handbook for acoustic ecology, Leonardo privacy and security concerns, ICT Express 9 (2023) 13 (1980) 83. 571–588. URL: https://www.sciencedirect.com/ [73] R. N. Bracewell, R. N. Bracewell, The Fourier trans- science/article/pii/S2405959523000243. doi:https: form and its applications, volume 31999, McGraw- //doi.org/10.1016/j.icte.2023.02.007. Hill New York, 1986. 30 Luca Corvitto et al. CEUR Workshop Proceedings 19–31 [74] C. Cortes, V. Vapnik, Support-vector networks, [87] M. S. Piotr Dabkowski, Eleven labs, 2023. URL: https: Machine learning 20 (1995) 273–297. //elevenlabs.io/voice-lab. [75] K. He, X. Zhang, S. Ren, J. Sun, Deep resid- ual learning for image recognition, 2015. arXiv:1512.03385. [76] T. W. Clément Delangue, Julien Chaumond, Wav2vec2, 2022. URL: https://huggingface.co/docs/ transformers/model_doc/wav2vec2. [77] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, E. Duch- esnay, Scikit-learn: Machine learning in Python, Journal of Machine Learning Research 12 (2011) 2825–2830. [78] L. M. Justice, W.-Y. Ahn, J. A. Logan, Identifying children with clinical language disorder: an appli- cation of machine-learning classification, Journal of learning disabilities 52 (2019) 351–365. [79] J. Gaspers, K. Thiele, P. Cimiano, A. Foltz, P. Sten- neken, M. Tscherepanow, An evaluation of mea- sures to dissociate language and communication disorders from healthy controls using machine learning techniques, in: Proceedings of the 2nd acm sighit international health informatics sympo- sium, 2012, pp. 209–218. [80] C. Kanimozhiselvi, S. Santhiya, Communication disorder identification from recorded speech using machine learning assisted mobile application, in: 2021 Third International Conference on Intelligent Communication Technologies and Virtual Mobile Networks (ICICV), IEEE, 2021, pp. 789–793. [81] A. Caliskan, J. J. Bryson, A. Narayanan, Semantics derived automatically from language corpora con- tain human-like biases, Science 356 (2017) 183–186. [82] D. Rozado, The political biases of ChatGPT, Soc. Sci. (Basel) 12 (2023) 148. [83] M. A. Gianfrancesco, S. Tamang, J. Yazdany, G. Schmajuk, Potential biases in machine learn- ing algorithms using electronic health record data, JAMA Intern. Med. 178 (2018) 1544–1547. [84] H. Ibrahim, X. Liu, N. Zariffa, A. D. Mor- ris, A. K. Denniston, Health data poverty: an assailable barrier to equitable digital health care, The Lancet Digital Health 3 (2021) e260– e265. URL: https://www.sciencedirect.com/science/ article/pii/S2589750020303174. doi:https://doi. org/10.1016/S2589-7500(20)30317-4. [85] P. Grill, J. Tučková, Speech databases of typical children and children with SLI, PLoS One 11 (2016) e0150365. [86] A. McAllister, P. Sjölander, Children’s voice and voice disorders, in: Seminars in speech and lan- guage, volume 34, Thieme Medical Publishers, 2013, pp. 071–079. 31