Neural models of speech processing comprise the modeling of speech production and speech perception/comprehension. Production models start with the specification of a verbal intention at the semantic or concept level, generate lemmata and phonological forms (cognitive-linguistic model part). These models subsequently initiate the motor execution processes including articulatory movement generation, acoustic signal generation, and sensory feedback signal generation (sensorimotor model part). Well-known neural models representing the sensorimotor part of speech production have been developed by Dell (1986) and Dell et al. (2007, 2013; spreading activation model), Roelofs (1992, 1997, 2014; WEAVER model), and Levelt et al. (1999; word production model). Well-known neural models representing the sensorimotor part have been developed by Guenther (2006, 2016; DIVA model), Guenther et al. (2006; DIVA model), and Bohland et al. (2010; GODIVA model). A biologically inspired feedback-aware speech task control approach has been introduced by Parrell et al. (2019; FACTS) and a spiking neuron model covering the linguistic and sensorimotor part has been developed by Kröger et al. (2012, 2016, 2020, 2022; ACT model). All these neural models are concrete, quantitatively implemented, and checked by computer-simulating realistic communication scenarios.

A comprehensive neurobiologically motivated but still not computer-implemented model of speech perception and comprehension is introduced by Hickok and Poeppel (2007, 2016). This model comprises modules for spectro-temporal analysis of incoming acoustic speech signals, for phonological processing and then splits in a ventral processing stream including lexical, semantic, and grammatical processing and in a dorsal stream for further auditory, somatosensory, and motor processing.

Combined production-perception models (speech processing models) are needed if the simulation of speech learning (i.e., modeling of speech acquisition) is of interest (developmental neural models; see Warlaumont and Finnegan, 2016; Kröger et al., 2022). During the babbling phase—which appears in the first year of lifetime—first sensorimotor relations establish (mainly auditory-to-motor relations) and later during the imitation phase, in which the toddler imitates speech items produced by caretakers, the mental lexicon and the grammatical repertoire of the target language is learned and stored. These speech acquisition phases are needed to be simulated in a realistic neural speech processing model. Moreover, complete neural speech processing models are needed for the simulation of speech communication scenarios, i.e., scenarios, comprising capabilities like listening to and speaking with a communication partner. It should be kept in mind that all medical screenings in case of diagnosis of speech disorders include such communication scenarios between test supervisor (communication partner) and patient (model).

It will be shown that the neural models reviewed here are able to unfold the complex associations between neural dysfunctions and symptoms of disordered speech in case of neurogenic speech and language disorders. Thus, models can help to refine the definition of speech and language disorders because an underlying neural dysfunction, which is the basis for the definition of a speech disorder, can be clearly defined in a neural model while a lesioned brain region of patients (i.e., anatomical information) which probably is located by imaging techniques in a patient (e.g., Crinion et al., 2013) does not necessarily point in a one-to-one relation on a specific neural dysfunction (i.e., functional information) nor on a specific speech and language disorder. While functional information is directly defined as inserted distortion to the model in a specific neural subnetwork, this information needs to be extracted indirectly from behavioral data by asking patients to perform specific speech and language tests in order to collect data of relevant speech errors from these screening scenarios.

In this paper we will concentrate on four types of neurogenic speech and language disorders, i.e., on aphasia, dysarthria, apraxia of speech, and neurogenic stuttering. Aphasia can be defined as a disorder resulting from neural dysfunctions arising in the cognitive-linguistic part of the speech processing network. Aphasia can affect the activation of a word at the lexical level even if motor processes are intact or can affect the comprehension of a word even if auditory perception is intact (e.g., Roelofs, 2014). Dysarthria and apraxia of speech result from neural dysfunctions in the sensorimotor part of the brain including the peripheral motor neuron system. All types of dysarthria reflect functional deficits appearing during motor execution even in case of fully functional articulatory organs (Kearney and Guenther, 2019). Apraxia of speech reflect deficits in motor planning and motor programming (Van der Merwe, 2021). Neurogenic stuttering reflects deficits in the initiation of execution of motor programs (Chang and Guenther, 2020).

Symptoms of speech and language disorders typically appear in communication situations and can be evoked in speech tasks like picture naming, narration tasks, word, non-word (logatome), or syllable repetition tasks, or in word or sentence comprehension tasks. For all types of neurogenic speech and language disorders, diagnosis procedures (also called screenings) comprise a batterie of tests and most of these tests are speech mediated tasks (i.e., the supervisor instructs the patient verbally, gives test items verbally and the patient answers verbally; a non-speech mediated task would be picture pointing like in the Token Test; here even the target words could be presented non-verbally, for example as written text). Well-known and widely used screenings in case of suspected aphasia are e.g., the Token Test (De Renzi and Vignolo, 1962; De Renzi and Faglioni, 1978), the Frenchay Aphasia Screening Test (FAST, Enderby et al., 1987), the Acute Aphasia Screening Protocol (Crary et al., 1989), the Aachen Aphasia Bedside Test (Biniek et al., 1992), and the Bedside Western Aphasia Battery (Kertesz, 2006). These screenings typically (i) assess comprehension, e.g., by pointing on objects on cards portraying a scene and/or geometric shapes, by executing simple movements based on instructions given by the test supervisor, (ii) assess expression, e.g., by describing a picture, by repeating words, or by naming of objects displayed on pictures, (iii) assess reading capabilities by reading words or short texts, and (iv) assess writing capabilities by writing words or a short text which describes a scene displayed on pictures.

Screenings for detecting dysarthria or apraxia of speech are often combined with screenings for differential diagnosis together with suspected aphasia and are sometimes subsumed as screenings for neurological communicative disorders (e.g., Araki et al., 2021) or as screenings for differential diagnosis of different types of neurogenic speech disorders (e.g., Allison et al., 2020). These screenings include verbal-linguistic sections (e.g., word and nonword repetition, object naming, word writing, dictation) and articulatory sections including non-speech tasks like oral movement analysis and tasks like diadochokinesis, i.e., repetition of syllable sequences like [badaga] or [pataka] as often as possible and as fast as possible. Apraxia of speech screenings as well include verbal-linguistic tests and articulatory tests like word and non-word repetition, sentence production, and phonological awareness tests. Here, the analysis of speech items which are uttered by patients in addition comprises phonetic transcriptions to identify prosodic and segmental errors (Ballard et al., 2016; Allison et al., 2020).

Treatments for all neurogenic speech disorders mainly comprise practice for improving speaking capabilities in case of sentence and word production. During ongoing treatment, the training concentrates on speech items with increasing length and complexity. Lexical learning strategies for the association of phonological word form and meaning aim to widen the vocabulary of patients in case of patients suffering from different forms of aphasia (Tippett et al., 2015). Practice for syllable production to learn the pronunciation of different speech sounds in typical speech-like environments and in combination with other speech sounds within a syllable is focused on in case of treatments for patients suffering from apraxia of speech (Ballard et al., 2000). In case of dysarthria in addition detailed advises are given for increasing or reducing speaking rate and speaking intelligibility or for increasing speech and non-speech motor capabilities for the neuromuscular system of several articulators including respiration and phonation (Palmer and Enderby, 2007).

In this paper a comprehensive sketch of a computer-implementable speech processing model is introduced (Figure 1 and Kröger et al., 2022). Figure 1 illustrates that models of speech processing can be divided in functional subnetworks or modules which specify functional locations of parts of the neural network and that each of these subnetworks or modules can be associated with specific cortical and subcortical brain regions. The model sketch presented in Figure 1 comprises a cognitive-linguistic model part for which the associations of subnetworks or modules to brain regions are outlined by Roelofs (2014) and a sensorimotor model part for which these relations are outlined by Guenther (2006), Bohland et al. (2010), and Kearney and Guenther (2019).

This model sketch is mainly based on two computer-implemented model approaches, i.e., on the DIVA/GODIVA approach for sensorimotor control of speech production (Guenther, 2006, 2016; Guenther et al., 2006; Bohland et al., 2010) and on the WEAVER approach for word-form encoding (Roelofs, 1992, 1997, 2014). The WEAVER model (as published by Roelofs, 2014) reflects the cognitive-linguistic part of the model sketch (Figure 1) and is based on the word production model published by Levelt et al. (1999). The DIVA/GODIVA model (as published by Miller and Guenther, 2021) reflects the sensorimotor part of the model sketch (for the differences between DIVA and GODIVA see below in this section). WEAVER as well as DIVA/GODIVA are implemented using second generation neural networks (see Appendix A). The network model developed by Kröger et al. (2016, 2020, 2022) aims for a complete representation of a speech processing network as given in Figure 1 including cognitive-linguistic and sensorimotor processing (see below) and uses a third generation neural network approach (see Appendices A–C).

The model sketch (Figure 1) comprises a mental lexicon and a mental syllabary as central knowledge and skill repositories within the cognitive-linguistic and sensorimotor part of the model (Figure 1A). The core of the mental lexicon—storing and processing cognitive speech states (concepts, lemmata, and phonological forms)—is located in the temporal lobe. Its brain locations overlap with the network part representing the auditory input states of syllables within the mental syllabary. Phonological representations as output of the mental lexicon and input for the sensorimotor part of the speech production part of the speech processing model sketch are located in the posterior part of the frontal lobe near the syllable initiation module of the speech production network.

Motor program states and somatosensory states of syllables as part of the mental syllabary are stored in the inferior parts of the frontal and parietal lobe (the transformation of phonological states into motor plans is described in detail by Bohland et al., 2010, i.e., within the GODIVA model, and this transformation is in accordance with the model concept given by Kröger et al., 2022; the processing using the mental syllabary is described in detail by DIVA model, see, e.g., Guenther et al., 2006). Speech perception is mainly located in the superior and posterior part of the temporal lobe and comprehension leads to lexical activations in the anterior part of the temporal lobe. The cortico-cortical feedback loop including basal ganglia and thalamus (action control loop) can be activated from many cortical regions and feeds neural activations back to different parts of the cortical speech processing neural network, while the cortico-cortical control loop including cerebellum via pons (motor feedback loop) is activated mainly by the production part of the sensorimotor network and feeds back its activations to this region for activating the learned auditory and somatosensory states for currently produced syllables as well as for activating motor program execution.

While imaging and lesion studies support the strong correlation between structural (anatomical) brain locations and functional aspects (functional modules) of the neural network (e.g., Batista-García-Ramó and Fernández-Verdecia, 2018; Litwińczuk et al., 2022), this does not implicate that there exists no close neighborhood or even spatial overlap of functional modules in several regions of the brain. Thus, functional deficits appearing in modules or sub-networks of the neural speech processing network cannot always be easily associated with a specific localization of dysfunctional (e.g., damaged) regions within the brain. Moreover, in the case of developmental speech and language disorders (i.e., delay of learning and storing data within the speech processing neural network; difficulties in learning), in case of speech and language disorders which result from neurodegenerative diseases, or in case of aging which may lead to (slow and limited) degeneration of the neural network, imaging studies may not indicate any specific anatomic regions or structural abnormalities which directly uncover an underlying neural deficit which probably is responsible for the occurring speech or language disorder. In all these cases specific screenings are needed in order to collect relevant behavioral data for diagnosing a speech and language disorder correctly.

It is possible to insert neural dysfunctions of any type and severity to any functional subnetwork or module (i.e., functional location) of a speech processing neural network model. Thus, a modeled neural dysfunction can be specified with respect to functional location, severity, and neural type. The functional location (i.e., a specific module or subnetwork in the model, which is affected) is correlated to a lesioned brain regions which can be identified on basis of functional imaging data, but in many cases, the identified brain areas hosting a specific sub-network, module, or buffer of the speech processing neural network are relatively broad (e.g., Golfinopoulos et al., 2010; Kearney and Guenther, 2019). The severity of a dysfunction is defined as the percentage of non-functioning neurons or of non-functioning neural connections within a module or sub-network of the neural model (e.g., Roelofs, 2014; Kröger et al., 2020). The neural type of a dysfunction separates dysfunctions of neurons (cells, specifically cell body), dysfunctions of synapses (synaptic connections), and dysfunctions of connecting pathways (axons, dendrites) within the modeled neural network (e.g., Roelofs, 2014). Moreover, some models are capable of varying concentrations of neurotransmitters like dopamine level in specific modules or subnetworks (e.g., in striatum of basal ganglia, Civier et al., 2013; Senft et al., 2016, 2018). In these models, an abnormal concentration (too low or too high) of a transmitter substance can be introduced for instantiating a further type of neural dysfunction.

As already stated above, the linguistic-cognitive part of the model sketch given in Figure 1 is mainly based on the WEAVER model and the sensorimotor part is based on the DIVA/GODIVA model.

WEAVER (Roelofs, 2014) is a second generation or node-and-link neural network (see Appendix A) consisting of seven node layers (or simply layers), separating concept level, lemma level, phonological form level, and syllable motor program level. Two input-/output-layers are labeled as lexical input and output layers and phonological forms are separated in input- and output-layers as well, while the lemma and concept level do not separate input and output forms. The syllable motor program layer in WEAVER is comparable to a motor plan level in our model sketch (Figure 1). Links are building up neural pathways for connecting different layers of the model, i.e., the layers representing concept, lemma, lexical in-/output, phoneme in-/output and motor plan level. From the functional viewpoint of neural processing these inter-layer neural connections or inter-layer links can be labeled also as neural mappings while phonemes, lemmata, and concepts are represented as neural states. Each state is represented in WEAVER by a specific node in a neural layer. Thus, this network type uses a local representation of states. The performance of the WEAVER network for simulating different types of aphasias and the temporal specification of increasing/decreasing node activation is discussed in Section 3 and in Section 5.

The DIVA/GODIVA approach differentiates 10 layers, also called neural maps in the context of this modeling approach (Guenther et al., 2006; Bohland et al., 2010; Miller and Guenther, 2021). These neural maps and their hypothetical location in the brain are discussed in Section 3 of this paper. The neural maps (i.e., initiation map, speech sound map, auditory target, state, and error map, somatosensory target, state, and error map, articulator map, and feedback control map) and the cortical mappings connecting these neural maps are displayed in Figure 1A. Here, the speech sound map is labeled motor plan map (or motor plan buffer), the feedback map is part of the mental syllabary buffer, and the articulator map is part of the motor execution buffer. These renaming of map labels in Figure 1 results from the separation of motor planning and motor programming and on quantifying motor plans and programs with respect to the concept of speech actions or articulator gestures (see Kröger et al., 2022). The functioning of the DIVA/GODIVA model and the modeling of speech disorders is discussed in Section 3 and in Section 5.

Moreover, the model sketch presented in this paper is in accordance as well with three further computational neural network models cited in this review paper (see below).

So far, the cognitive-linguistic part as well as the production-side of the sensorimotor part of the model sketch (Figure 1) are computer-implemented now by using a spiking neuron approach (Kröger et al., 2016, 2020, 2022; see also Appendix C). The feedback loop of the sensorimotor part of the model sketch has been implemented beside DIVA in a spatio-temporal activation averaging model (STAA model or second generation neural network model, Kröger et al., 2014; Kröger and Cao, 2015; Kröger and Bekolay, 2019, p. 133ff, while spiking neuron or spiking neural networks i.e., SNN's, are also called third generation neural network approaches, see Maass, 1997).

Computer-implemented neural network models are simulating neural functionality. These models clearly define subnetworks which are responsible for specific functional subtasks, e.g., for selecting and activating a concept, lemma, or phonological form, or for activating a stored syllable motor plan etc. Imaging techniques allow to specify exactly those brain regions which are activated if a specific functional task is performed and thus allow to associate neural functionality and brain areas. Indefrey and Levelt (2004) and Roelofs (2014) assume that concepts which are stored in the mental lexicon are represented in anterior-ventral temporal cortex, lemmas in the mid-section of the left middle temporal gyrus, input and output lexical forms of lemmas as well as input phonemes in left posterior superior and middle temporal gyrus (Wernicke's area), while output phonemes are stored in left posterior inferior frontal gyrus (Broca's area). Syllable motor representations which are stored in the mental syllabary are represented in ventral precentral gyrus. Inter-lobe neural associations appear especially between phonological input and output forms located in part in the temporal lobe and in part in the frontal lobe (see Figure 1B). These associations are structurally realized by left arcuate fasciculus and uncinate fasciculus.

Guenther (2006), Guenther et al. (2006), Golfinopoulos et al. (2010), and Kearney and Guenther (2019) assume that the initiation map, which activates motor plans and motor programs and thus starts syllable execution as postulated in the DIVA and GODIVA models (Guenther, 2006; Bohland et al., 2010), is located in the supplementary motor area, on the medial wall of the frontal cortex. The speech sound map (a term used in DIVA and GODIVA models and represented by mental syllabary in our model sketch, Figure 1A) is assumed to be located in left ventral premotor cortex, i.e., in the ventral precentral gyrus and in the surrounding portions of posterior inferior frontal gyrus and of the anterior insula.

The articulation map (execution map in model sketch, Figure 1A) which directly activates motor neurons controlling the movements of the speech articulators is located within the ventral motor cortex (primary motor cortex). Neural buffers hosting the auditory target, state, and error maps are located within the ventral auditory cortex (temporal lobe), and those hosting the somatosensory target, state, and error maps are located in the ventral somatosensory cortex (parietal lobe). The detailed organization and functioning of the feedforward and feedback motor control system within the sensorimotor part of the model sketch using these maps is described by Guenther (2006) and Kearney and Guenther (2019) and in the context of our model sketch it is described by Kröger et al. (2020, 2022).

Two cortico-cortical feedback loops (action control loop and motor feedback loop) including basal ganglia, cerebellum, and thalamus are introduced in the model sketch (Figure 1A). The action control loop (see orange arrows, orange box, and orange text in Figures 1A, C) is responsible for all cognitive control processes needed for temporal sequencing of cognitive and sensorimotor processes in each situation, e.g., paying attention to specific incoming sensory information, deciding how to react in a specific situation, and activating motor processes for reacting. In case of a speech task like picture naming this can be the sequence of visual perception (activation of visual state), word recognition (activation of a concept from mental lexicon), and word production (phonological form activation, Kröger et al., 2016, 2020). Moreover, action control as well comprises motor planning in form of motor plan and motor program activation and motor execution (see planning and motor loop in Bohland et al., 2010 and in Miller and Guenther, 2021). This control loop starts and ends in different areas of the neocortex and includes the basal ganglia and thalamus in its center (see Figures 1B, C, solid orange lines). A second feedback loop, here called motor feedback loop (called cerebellar loop in Kearney and Guenther, 2019; green arrows and boxes in Figure 1A and green lines and green text in Figure 1C) is responsible for activating feedback control processes comparing stored and current auditory and somatosensory states and thus in part acts on syllable execution as well. This loop is activated together with current target states at mental syllabary, comprises cerebellum and thalamus in its center, and allows correction of sensory feedback states as well as of motor program states (see Figures 1A, C, green lines).

While most approaches discussed above can be represented by functionally defined box-and-arrow models, i.e., can be represented by box-and-arrow plots in which the boxes or modules define partial functions of speech processing like semantic-to-lemma or lemma-to-phonological form transformation (see, e.g., Roelofs, 2014), the LICHTHEIM 2 approach (Ueno et al., 2011) can be represented by a neuroanatomically defined box-and-arrow model. Here, the architecture of the network and thus the modules of the network are defined with respect to the neuroanatomy of the brain, but it should be kept in mind that these brain regions are mainly defined based on knowledge gathered from functional imaging experiments, so that these regions as well can be separated on the basis of neurofunctionality.

The LICHTHEIM 2 model is based on a second generation neural network model and separates seven neural layers representing different cortical areas. The model (i) connects the auditory input layer (mid-superior temporal gyrus/sulcus, mSTG, mSTS) with the semantic layer (anterior STG/STS) by two intermediate or hidden layers within the temporal lobe via the ventral pathway, (ii) connects the semantic layer with the motor output layer (insular motor cortex) via two intermediate or hidden layers within the temporal and frontal lobe as a further part of the ventral pathway (main part of the ventral pathway), and (iii) connects the auditory input layer and motor output layer via one further intermediate or hidden layer within the parietal lobe (i.e., a hidden layer within the inferior supramarginal gyrus, inf SMG) by the dorsal pathway.

Within the functional part of our model sketch (Figure 1A) a part of the ventral route located in the temporal lobe and the further dorsal route (both routes as defined by Ueno et al., 2011) can be interpreted as the production pathway (our model sketch, Figure 1A) leading from semantic processing via lexical processing (from the concept buffer via lemma, to phonological form buffer), toward motor processing (pathway from motor via and program buffers toward motor execution). In the other (i.e., in the perceptual) direction (the perception pathway in our model sketch), the dorsal pathway (from frontal via parietal toward temporal lobe as defined by Ueno et al., 2011) can be interpreted as sensory feedback processing pathway including somatosensory and auditory state, target, and error buffers (Figure 1A), while the part of the ventral pathway located in the temporal lobe (as defined by Ueno et al., 2011) as well realizes perceptual lexical processing (i.e., comprehension).

In contrast to all other neurocomputational models and as stated above, the LICHTHEIM 2 approach does not specify the intermediate or hidden layers with respect to sensorimotor or linguistic functions (e.g., phonological, higher level auditory, or higher-level somatosensory representations) and thus can be interpreted as a basically neuroanatomical approach. Only the input- and output layers are defined concerning functions, i.e., the semantic layer for activation of word meanings, the acoustic input layer for activation of phonetic-acoustic features and the motor output layer for the activation of motor programs. Thus, the model can be interpreted as well as an early version of deep learning networks (for an overview on deep learning neural networks in speech processing see, e.g., Nassif et al., 2019; Roger et al., 2022). A further hidden layer neural network model for speech processing including the ventral visual pathway but omitting a part of the speech processing ventral pathway (connecting the anterior part of the temporal lobe directly with the inferior part of the frontal lobe) has been developed by Weems and Reggia (2006).

The main questions which will be answered in this chapter for each existing computer-implemented neural model is: Which speech and language disorder is intended to be simulated with this dysfunctional model? Which dysfunctions (type and location) are inserted in that model to simulate these speech and language disorders? Which simulation scenarios (screening tasks) were simulated using that dysfunctional model in order to simulate all typical symptoms of the targeted speech and language disorder?

Five main groups of computer-implemented neural models will be discussed here, i.e., WEAVER, Dell's spreading activation model, LICHTHEIM 2, DIVA/GODIVA and ACT. The WEAVER model of Roelofs (2014) as well as the spreading activation model of Dell et al. (2013) are second generation network models (see Appendix A) comprising semantic, lemma, and phoneme layers (comparable to neural maps in DIVA/GODIVA and to buffers in ACT) of nodes (comparable cells in DIVA/GODIVA and to neuron ensembles in ACT) connected by bidirectional inter-layer links or inter-layer neural pathways (representing neural connections). While the terms layers, nodes, links are defining second generation neural network approaches (using spatio-temporal activation averaging, see Kröger and Bekolay, 2019, p. 133ff and see Appendix A) the terms buffers, cells and synaptic connections are used in adaptive neural network approaches (DIVA: Guenther, 2006, 2016; Guenther et al., 2006, GODIVA: Bohland et al., 2010) and in third generation neural network approaches also called spiking neuron models (Kröger et al., 2016, 2020, 2022; Senft et al., 2016, 2018; Stille et al., 2020; also labeled as ACT model, see Kröger et al., 2012; and see Appendices A, B).

The WEAVER model (Roelofs, 2014) can simulate cognitive-linguistic production and perception/comprehension processes and is able to generate typical speech symptoms appearing in different forms of aphasia, i.e., Broca's, Wernicke's, conduction, transcortical motor, transcortical sensory, and mixed transcortical aphasia. These symptoms comprise production or comprehension of a wrong word, a complete failure of word production, or in case of nonwords (meaningless syllables or syllable sequences) the production of no or of a wrong or reduced sequence of speech sounds. These symptoms typically appear in question-answering scenarios and are tried to be evoked in medical screenings (questions by test supervisor; answers by patient) comprising picture naming, word repetition, word comprehension, or logatome (i.e., nonword) repetition tasks (see introduction for the definition of these tasks and see Table 4). The LICHTHEIM 2 approach (Ueno et al., 2011) can simulate different types of conduction aphasia, Broca- and Wernicke aphasia by simulating the same types of tests, i.e., naming, word comprehension and word and logatome repetition. The brain lesions inserted in this model are not primarily functional but defined from cortical locations but can be interpreted in a functional way as is explained already in Section 3.

To evoke these symptoms by simulation, two neural types of dysfunctions can be chosen in WEAVER (ibid.). (i) Reduction in strength of neural activation appearing in the nodes because of a specific percentage of inactive or dead neurons or (ii) reduction in strength of neural activation forwarded in synaptic connections between neurons which results from a specific percentage of inactive or dead synaptic connections or links. These types of model dysfunctions can be inserted in neuron buffers at concept, lemma, or phonological form levels on the perception or production pathway or can be inserted in the neural connections between these buffers within the production or perception pathway (see Figure 1). The severity of a dysfunction is modeled in WEAVER by (i) the parameter decay rate affecting the nodes (i.e., the model layers), and by (ii) the parameter connection weight affecting the links (i.e., the connections between layers). Thus, the stronger the decay rate or the lower the connection weight the higher the number of damaged nodes or links and the higher the rate of symptoms produced in simulated speech tasks.

A comprehensive description of typical simulations for generating symptoms in different forms of aphasia using the WEAVER model is given by Roelofs (2014). The neural dysfunctions corresponding to different forms of aphasia are inserted at the phonological, the lemma, and the concept in- and output layers in form of increase in decay rate which models an increasing number of dysfunctional neurons within these layers. Furthermore, neural dysfunctions are inserted in form of decrease in connection weight for modeling an increasing number of dysfunctional neural connections between layers (see chapter 2 of this paper and see ibid., p. 37f). Three types of tests were simulated. (i) Word production is simulated by introducing neural activation at the phonological input layer representing a phonological word form of a specific target word and by evaluating whether the correct activation appears at the syllable nodes below the phonological output layer (i.e., motor plan nodes in Figure 1). (ii) Word comprehension is simulated in the same way concerning target word activation but here, the activation at the concept level layer is evaluated. (iii) Logatome repetition is simulated in the same way as word production but here instead of target words target syllables are activated for which no corresponding lemma and concept exists (i.e., phonologically well-formed syllables without word meaning in the target language). If logatomes are not part of the model vocabulary, it is possible to simulate logatome repletion by using words or syllables of the target language if the neural connection between phonological form level and lemma level is interrupted in the model at the perception side in order to avoid a coactivation of word forms and concepts.

Simulation results of these types of tests are interpreted as an error if no neural activation arises at the concept level (comprehension test) or at the motor plan level (word production and logatome repetition test) or if the occurring neural activation pattern represents a wrong word. The results from this interpretation of simulated errors indicate for all types of aphasia that the error rate increases with increasing neural damage of layers or neural connections at least for one of the three tests. Thus, a strong error rate appears for the word production and repetition test in case of neural damage within the phonological output layer (Broca aphasia), a strong increase in error rate appears for word comprehension and repetition in case of neural damage within the phonological input layer (Wernicke aphasia), a strong increase in error rate appears for logatome repetition in case of neural damage within the neural pathway between input and output phonological layer (conduction aphasia), a strong increase in error rate appears for word production only in case of neural damage within the neural pathway between lemma and output phonological layer (transcortical motor aphasia), a strong increase in error rate appears for word comprehension only in case of neural damage within the neural pathway between input phonological and lemma layer (transcortical sensory aphasia), and a strong increase in error rate appears for word production and comprehension in case of neural damage within the neural pathway between lemma buffers and concept layers within input or perception and within production or output pathway (mixed aphasia; see ibid, Figure 2 on p. 38).

Dell (1986) introduces a weight-decay approach implemented as dysfunction-inserting approach in his spreading activation model which represents a network-wide reduction in connection weight (weight parameter affecting links and thus the connections between layers) combined with a decrease or increase in activation-decay rate (decay parameter affecting activation of nodes within a layer). In later versions of the spreading activation model the weight parameter is split into separate semantic (s-weight) and phonological (p-weight) weights representing different functional locations of dysfunctions (between semantic and word layer and between word and phoneme layer), while the activation decay in nodes cannot be changed. The continuity thesis (Schwartz et al., 2006, p. 232) implies that an increase/decrease, i.e., the strength of each of these parameters changes the model performance from normal to random (i.e., toward incorrect or abnormal behavior) while the quotient of strength of these parameters, i.e., the degree of dominance of one of these parameters, characterizes the type of disorder. An auditory input layer and a further weight parameter (nl-weight) is introduced by Dell et al. (2013) which describes dysfunctions between the auditory input layer (added in this new model variant) and phoneme layer and which in addition allows the modeling of auditory input disorders. The tasks simulated by spreading activation models were (picture) naming by inserting primary or input neural activation at the semantic layer, and word and nonword repetition by inserting neural activation at the auditory layer of the model (see Table 4). Furthermore, Dell et al. (2013) introduces a concept which allows an association of cortical locations of brain lesions and model dysfunctions by associating model and patient behavioral data. The simulation results are comparable with those generated by the WEAVER model (Roelofs, 2014).

In the LICHTHEIM 2 model (Ueno et al., 2011) brain lesions are defined from a neuroanatomical viewpoint as white matter damage and gray matter damage. Gray matter damage is modeled here as an insertion of white noise to the activation pattern of nodes within a specific network layer (damage within a specific layer) while white matter damage is modeled as a partial removal of neural links of a neural pathway interconnecting two neighboring neural layers of the network model (damage of a specific neural pathway). Both types of damage are applied in combination with increasing severity leading to an increasing number of speech errors appearing in the simulated naming, word comprehension, and word and logatome repetition tasks.

Because the LICHTHEIM 2 model is capable of modeling lexical relearning by applying a set of training stimuli in form of auditory and related motor activation patterns to the auditory and motor layer and by adjusting synaptic connections of the hidden layers along the lexical route within the temporal lobe, this model is capable of simulating post-stroke neural recovery via the ventral route if the dorsal route is damaged (see Stefaniak et al., 2020, p. 47ff). This relearning (readjustment of synaptic link weights) applied to the model can be interpreted as a post-stroke learning or recovery.

The adaptive neural production models DIVA and GODIVA developed by Guenther (2006, 2016), Guenther et al. (2006), and Bohland et al. (2010) simulate neural processes of speech learning (early phases of speech acquisition, i.e., babbling and imitation) and neural processes of feedback-controlled speech production (sensorimotor part of the speech production model; feedforward and feedback control). These models comprise motor planning (i.e., selection of executable chunks at the phonological level), motor program selection, and motor program execution. The DIVA/GODIVA model components (modules or subnetworks) are associated with specific cortical as well as subcortical brain regions (Kearney and Guenther, 2019, and see above, chapter “anatomical locations”) and several modules or subnetworks can be identified in the DIVA/GODIVA models which cause symptoms of dysarthria or apraxia of speech (Kearney and Guenther, 2019, pp. 11ff; Miller and Guenther, 2021, pp. 432ff, and see above, chapter “disorders”). A concrete simulation study using the GODIVA model has been performed for simulation neurogenic stuttering (Civier et al., 2013, Table 3). Model parameters characterizing neural dysfunctions were (i) number of defective cells within distinct cortical or subcortical modules, (ii) number of defective neural connection weights between cells of distinct modules of the model, and (iii) change in dopamine level within striatal component of the modeled basal ganglia module of the model. Typical symptoms which can be simulated using this model by performing a word production task are (i) slower initiation of execution of a motor program (ibid., p. 272), leading to prolongation of preceding syllables as well as to silent blocks.

The spiking neuron model developed by Kröger et al. (2016, 2020, 2022), also called ACT model (Kröger et al., 2012) is capable of simulating speech errors as produced by normal speaking subjects (picture naming task without and with auditory distractor signals, Kröger et al., 2016, see also Table 4), is capable of simulating speech errors produced by speakers suffering from different forms of aphasia (word comprehension tasks and word and nonword production tasks, Kröger et al., 2020, p. 13f) and by subjects suffering from developmental (neurogenic) speech deficits concerning lexical access (picture naming tasks with auditorily presented phonological or semantic cues, Kröger et al., 2020, p. 14f), and is capable of simulating symptoms of neurogenic stuttering which result from changes in dopamine level within the basal ganglia (syllable repetition tasks, Senft et al., 2016).

In case of simulating different forms of aphasia the same types of simulations are used here for ACT as described above for the WEAVER model, i.e., simulations of word production, word comprehension and of logatome repetition. The same buffers and neural pathways are disturbed in the ACT model as already described above for the WEAVER model. But in the ACT model we are able to directly deactivate a specific number of neurons in the phonological form, lemma, and concept buffers. And in case of the neural pathways we are apbel to directly deactivate of a specific number of neurons within the associative memories because in ACT we have a direct modeling of neurons within buffers and within associative memories, while in WEAVER the neuron modeling is more indirect by using nodes and links (for the definition of nodes, links, neurons, neuron buffers and associative memories see Kröger and Bekolay, 2019, p.133ff).

Changes in dopamine level were also induced in the GODIVA model (Civier et al., 2013) which also resulted in typical symptoms of neurogenic stuttering, i.e., silent blocks and prolongation of syllables. In the ACT model used by Senft et al. (2016, 2018) changes of dopamine level (here reduction of dopamine level) led to symptoms of stuttering like omission of syllables, like errors in syllable ordering, as well as to repetitions of the same syllable. The task used here was repetition of a pre-learned sequence of syllables like [ba-da-ga] (this kind of task is also called diadochokinesis).

Simulations of a word repetition task performed with the normal model (healthy subject in the ACT model, Kröger et al., 2016) leads to speech errors (wrong word is produced or no word is produced) if the production process is distracted by placing perceptual events like distractor words during the production process of a target word within a picture naming task. Distractor words are most effective if they are semantically and/or phonologically similar with the target word, presented by the picture. An interesting result of this simulation study is that even the normal picture naming task executed in the ACT model without inserting any neural dysfunction produces a low rate of speech errors as it is the case for normal speakers in normal conversation or reading scenarios. This error rate increases dramatically in case of word repetition tasks including distractor words.

Simulation of picture naming in case of inserting model dysfunctions concerning neurons (rate of ablated neurons within a model buffer) and concerning neural connections (rate of ablated neurons in an association memory, see Kröger et al., 2020) for simulating subtypes of aphasia leads to reduction of correct word productions in picture naming tasks (Table 3). Errors appearing here are the production of wrong or of no words. Same holds for the non-word repetition task. The repetition of the correct syllable sequence goes down, and errors appear like production of a wrong syllable sequence or no syllable production at all. Comparable results appear for the word comprehension task. The comprehension rate becomes low and beside correct word meanings more and more wrong meanings become activated at the semantic level of the neural model or no item is activated at that level if the severity of the neural dysfunction (rate of ablated neurons) is increased in the model (Kröger et al., 2022).

Stille et al. (2020) was able to show that phonological or semantic cues can help to increase the rate of correct word productions in a picture naming task if a target word is not produced in a first trail (the model is not able to activate the correct word at the semantic or phonological level) and if cues are given by the environment (by a communication partner) by using the ACT model. Here, neural dysfunctions were inserted in the neural model at the semantic level of the production pathway in order to model lexical access problems.

In general, the simulations performed using the ACT approach (Kröger et al., 2020; Stille et al., 2020) suggest that the “punctual” model dysfunctions implemented in case of modeling different types of speech disorders lead to a variety of different speech symptoms like (i) phonological distortions if a wrong but phonologically similar syllable or word is activated, (ii) to a drop out of a word production or word comprehension if the activation of an item at the phonological, lemma, or concept level is too low, and (iii) wrong word production or wrong word comprehension if further (non-similar) items are co-activated at the phonological, lemma, or concept level.

Beside a detailed neurobiologically inspired architecture, a neural model of speech processing needs to be able to simulate different communication scenarios as they typically appear in medical screenings (the modeled speech tasks are already mentioned in chapter “Simulation of symptoms” in this paper). Thus, the neural model needs to be able to react on an auditory or visual input (stimuli) and the model should include initiation processes for activating the production process at the cognitive-linguistic level (semantic, lemma, and phonological level) by these stimuli and to further activate motor plans, motor programs, and motor execution as well as to activate the perception and comprehension process. In addition, neural models should include halt or correction procedures during an ongoing production process (see generation of an error signal by comparing intended and produced sensorimotor signals via sensorimotor feedback loop; Kröger et al., 2016). These model features mentioned above are available in the DIVA/GODIVA model (Guenther et al., 2006; Bohland et al., 2010), in the ACT model (Kröger et al., 2012, 2016, 2020, 2022), and in part in the WEAVER model of Roelofs (1992, 1997, 2014) as well as in the spreading activation model of Dell (1986) and Dell et al. (2007, 2013).

The simulation of communication scenarios like those appearing in medical screenings between test supervisor and patient requires (i) an always active perceptive input channel even during ongoing production processes, (ii) the modeling of temporal aspects for forwarding neural activation patterns along the perception and production pathways, e.g., the modeling of concept to lemma to phonological form conversion, (iii) the modeling of cortical and sensorimotor action control (modeling of cortico-cortical feedback loop including basal ganglia and thalamus) for initiating specific cognitive and/or motor action is dominant at specific points in time, and (iv) the modeling of motor feedback (modeling of cortico-cortical feedback loop including cerebellum and thalamus) in order to simulate online control for all motor actions.

Especially the action control component is important in order to model different communication scenarios (i.e., different tasks as part of a medical screening) because different scenarios need the activation of different processing paths within the neural model. In the case of picture naming an external visual stimulus (e.g., a picture displaying a specific object) is initially processed by the visual object recognition module leading to a neural activation at the entry of the cognitive-linguistic module for representing that visually activated item at the semantic level of the perception pathway (Figure 1A). Because the patient (the model) is instructed to name the object in this task, this semantic activation is directly forwarded toward the production pathway (by skipping further cognitive processing) which results in a cascade of neural transformations from semantic state via lemma state to phonological form state and further from motor plan via motor program activation toward execution of the motor program by activating the neuromuscular system. Subsequently, online feedback procedures activate the auditory and somatosensory state of the produced speech item (sensory feedback state) which can be compared with the auditory and somatosensory expectations (learned and stored target states), activated earlier within the production process, which may lead to an online repair of a not well-articulated syllable or to a halt of articulation in case of a severe articulation error. In case of normal speech, in most cases the sensory expectations (target states stored in mental syllabary) match with the feedback signal produced during articulation (currently produced sensory feedback states) and no error signal is generated and thus the neural processing for generating an articulatory correction or a halt signal, which is realized by the action control loop (Kröger et al., 2016, 2020), needs not to be activated.

Word repetition tasks start at the auditory input level, i.e., activation of an auditory (input) state by a signal (stimulus) produced by the communication partner (external speaker, Figure 1A; test supervisor in a medical screening scenario). The resulting auditory state activated at the auditory input buffer activates an auditory state of mental syllabary (syllable level) or can be analyzed in smaller chunks leading to an activation of speech sound candidates (e.g., Guenther et al., 2006). In both cases this leads to an activation of phonological (input) states and subsequently leads to an activation of lemma candidates and to the activation of a semantic state within the perception/comprehension pathway (Figure 1A). The activated semantic state (concept buffer, see Figure 1A) directly activates the production pathway in the case of the word repetition task and subsequently a word candidate is selected within the production pathway and subsequently processed and its motor program is executed in the same way as described above for the picture naming task (e.g., Kröger et al., 2020).

In the case of non-word repetition (logatome repetition), i.e., repetition of a syllable sequence with no meaning in the target language (mother tongue or learned language), the neural activation at the phonological level does not lead to an activation of a lemma and of a concept stored in the mental lexicon. In this case the shortcut between phonological input and output phonological level is activated (dashed black line in Figure 1A; and see Kröger et al., 2020) leads to a direct activation of the phonological form within the production pathway of the model (phonological output buffer, Figure 1A) leading to further activations of motor plans, motor programs and subsequently to motor execution for the activated logatome or syllable sequence.

Word comprehension tasks activate the same neural pathway as already mentioned in case of the word repetition task. But here the processing already ends at the level of concept activation (activation of a meaning, e.g., Kröger et al., 2020). In case of a medical screening task for word comprehension, the target word is presented acoustically by the test supervisor and the patient is asked to point on one specific picture as part of a list of pictures to allow the test supervisor to see the word candidate, selected by the patient. This pointing procedure which is part of the scenario is not included in most models because neural activations can be directly accessed at all levels within the model (here at the concept level) which allows a direct monitoring of concept selection within the task by the patient (model) even without an explicit motor reaction.

In case of all these tasks (all these communication scenarios) described above the patient is already prepared or primed for giving a specific motor reaction, i.e., speaking by using the speech articulation apparatus or gesturing by using the arm-hand apparatus, if an input stimulus is seen or heard. Thus, in case of medical screenings the action control loop is already prepared for activating a specific sequence of cognitive and motor actions which is the consequence of a priming procedure, i.e., a consequence of preparing and instructing the patient or model for executing a specific task. Thus, even if a neural model does not include an action control loop (not including a model of basal ganglia and thalamus), the neural model can be shaped in a way for executing a specific task, i.e., by activating a specific input buffer (e.g., auditory input buffer or visual input buffer) which always leads to a chain of co-activations of further buffers in order to perform the neural processing required for executing a specific task. Thus, tasks can be performed also in case of the spreading activation model of Dell (1986) and Dell et al. (2007, 2013) and in case of the WEAVER model of Roelofs (1992, 1997, 2014) without an explicit modeling of action control.

A relatively complex medical screening task based on a complex communication scenario has been simulated by Stille et al. (2020) using the ACT model. Here, as part of a picture naming task—designed for quantifying lexical access mechanisms—semantic and phonological cues were provided auditorily (by the test supervisor) only in those cases, where a word is not directly produced by the patient (by the model) in a time interval of a few seconds. In these cases, a second and a third word production trial is started by providing acoustic phonological or semantic cues in parallel to the still available visual information. Thus, in case of this task, the action control loop allows word production directly by visual input (a picture representing the target word, for example a pic of a ball) an later the action control loop allows word production based on visual input but added by auditory input (test instructor gives a phonological cue like “the word starts with a [b]” or a semantic cue “the object can be thrown or kicked”). It has been shown by Stille et al. (2020) that in case of modeling a mental lexicon as it is acquired by children suffering from lexical access problems, lexical dysfunctions can be divided in “within level” and “between level dysfunctions,” i.e., in neural dysfunctions appearing within the neural buffers storing and ordering concept, lemma, or phonological form information with respect to semantic, grammatical, or phonological information and in neural dysfunctions appearing in the neural pathways between neural buffers for forwarding information from concept to lemma, or from lemma to phonological form buffers, i.e., in buffers representing the association of concepts to their lemmata as well as of lemmata to their phonological forms. Based on the simulations, Stille et al. (2020) found performance differences for lexical selection and activation processes if neural dysfunctions are located at the semantic and/or if neural dysfunction are located at the phonological level of the mental lexicon.

Neural modeling of speech processing allows to unfold the relations between location, type, and severity of a neural dysfunction inserted into a model and type and frequency of arising speech symptoms if simulations of specific communication scenarios (speech tasks) are performed using this model. In such a research endeavor the location of a dysfunction within a neural model is defined in a functional and not directly in an anatomically way. While a functional subnetwork or module of a neural model can be associated with an anatomic location in a direct way (see Roelofs, 2014 in case of the WAVER model; see Guenther, 2006; Guenther et al., 2006; Kearney and Guenther, 2019 in case of the DIVA/GODIVA models and Ueno et al., 2011 in case of LICHTHEIM 2 model) it in not always simple to identify a specific neural dysfunction on the basis of disruptions or damage appearing in a specific anatomical location within the central nervous system of a patient. Thus, it is not easy to associate brain lesions, disruptions, or abnormalities appearing in a specific anatomic location of the central nervous system (probably identified by neuroimaging methods applied to patients) and functional deficits in speech and language processing. For example, a brain lesion arising in the mid part of the temporal lobe may result in dysfunctions of different lexical submodules or buffers, for example lemma and concept buffer at the production as well as on the perception pathway. Or in case of a neurodegenerative abnormality of the basal ganglia we need to know in detail how this defect affects the cortico-striatal association network to differentiate dysfunctions (functional deficits) with respect to connections of the action control loop with the sensorimotor part or with the linguistic-cognitive part of the speech processing network. Thus, insertions of defined neural dysfunctions in computer-implemented quantitative neural models and its behavioral results generated (i.e., simulated) in speech tasks, allow to refine the definition of a speech or language disorder with respect to its etiology. It should be kept in mind that the LICHTHEIM 2 model here plays an intermediate role, because this model is not defined primarily in the neurofunctional domain (only the input and output layers are defined in a functional way) but refers to the neuroanatomical domain because the model separates different intermediate (or hidden) neural layers by specifying their neuroanatomical location but without specifying these hidden layers directly in a functional sense (e.g., for representing phonological, lemma, or concept forms).

Moreover, it should be kept in mind that a model could generate plausible simulation results even if the underlying neural mechanisms implemented in that model are not (strictly) similar with those appearing in humans. But in order to hold a high similarity of natural and simulated neural processes all neural models mentioned in this paper include basic as well as advanced neurofunctional knowledge gained from natural data as it is available from contemporary literature.

Furthermore, it needs to be mentioned that models are helpful to define functionality and to associate specific functionality appearing in specific brain regions with specific submodules of neural models, but it should be kept in mind that the hypothetical association of brain regions with submodules of neural models does not automatically strengthen the neurobiological reality of a model.

Nevertheless, the potential application of neural models in medical research could be manifold. (i) Neural models of speech processing allow to simulate medial screenings (tasks which are used for the diagnosis of speech and language disorders) by simulating the corresponding communication scenarios between patient and test supervisor. If location, type, and severity of an inserted dysfunction is varied in the model the simulation results could uncover the sensitivity of a screening task with respect to a specific neural dysfunction. Thus, simulations of screening tasks allow to estimate the effectiveness of that screening task to uncover a specific speech and language disorder. (ii) Because speech screening tasks usually are shaped or configured manually based on the experience of leading experts in the field of diagnosis and therapy of speech and language disorders, this information concerning effectiveness or sensitivity of a specific screening task could help to optimize screenings by varying all available task scenario parameters like type, number, or complexity of test items, number or repetition of trials, etc. (iii) Because of potential learning and familiarity effects and because of ethical reasons, a screening task can be undertaken only one time with a patient, while model simulations can be repeated as often as necessary by using the same computer-implemented model. Thus, a high number of patients is needed, all suffering from the same type and same severity of a speech or language disorder, to generate meaningful results for the optimization of a speech screening. Such a research endeavor is difficult to realize because of the need of a high number of well diagnosed patients for conducting such a research study, while neural models are capable to fulfill these demands more easily. A model is capable to repeat a screening task as often as needed and the location, type, and severity of the (inserted) neural dysfunction is clearly defined. This could result in a high reliability for the results concerning the association of location, type, and severity of a neural dysfunction and type and number of relevant speech symptoms. Moreover, no ethical conflicts appear in case of using computer models. (iv) All aspects discussed above are also applicable for the development of therapy scenarios. Because learning effects can be simulated in neural models as well, it is possible to quantify these learning effect and thus to quantify the efficiency of a therapy scenario for different types and severity levels of a specific speech and language disorder. Moreover, it is possible to vary all parameters of a therapy scenario (e.g., length of intervention, number and type of speech items trained in the therapy scenario, different designs concerning the increase in complexity of test items during a therapy, etc.) in order to find an optimally shaped treatment procedure (therapy scenario). But it needs to be stated here that all these ideas for simulation experiments have not been realized thus far. This disadvantage is mainly due to the current state of the art in neural modeling. Most models are currently used mainly for simulating typical behavioral effects like the production of striking symptoms, in order to exemplify the quality of modeling already reached in these days, but further model development and further simulations need to be done in order to simulate more complex screening or therapy scenarios in order to be able to deliver statistically significant results which allow to modify these scenarios toward more efficiency.

The detailedness of neural models of speech production is already on a relative high level in case of the cognitive-linguistic model part (Roelofs, 2014; Kröger et al., 2020) as well as in case of the sensorimotor model part (Guenther, 2006; Guenther et al., 2006; Bohland et al., 2010; Kröger et al., 2022). This allows a detailed modeling of speech and language disorders like aphasia with respect to all lexical aspects (production and comprehension deficits) as well as for apraxia of speech (production deficits with respect to planning and programming of syllables and syllable sequences). While the modeling of vocal tract geometries and vocal fold dynamics including all aerodynamic and acoustic relations is already on a high level as well (for a review see Kröger, 2022), there are still deficits in modeling the neuromuscular system and especially there is no overall consent concerning a control concept for neural activation at the level of the neuromuscular system directly controlling speech articulators (ibid.). This limits our current modeling endeavors for example concerning the simulation of articulatory consequences of dysarthric speech disorders especially in case of flaccid dysarthria and in case of spastic dysarthria (abnormal muscle tone). But due to the existence of already very detailed models of the basal ganglia and of the cortico-cortical control loop including basal ganglia and thalamus (Kröger et al., 2022), the modeling of other types of dysarthria is already possible but has not exemplified yet. In case of neurogenic stuttering the detailed implementation of this cortico-cortical action control loop already gave plausible simulation results for symptoms appearing in stuttering (see Civier et al., 2013).

Unfortunately, the neurobiologically based quantitative modeling of speech perception and speech comprehension is not as developed as it is the case for speech production. Specifically, there exist no comprehensive models which include the important concept of brain waves (gamma and theta waves, see Hickok and Poeppel, 2007; Ghitza, 2011; Ghitza et al., 2013) which is needed in order to model speech perception and speech comprehension realistically in a neurobiologically grounded way.

Because the microscopic functional level of natural neural networks—i.e., the cellular level of neurons and their functioning within subnetworks as well as within the whole network of speech processing—is not or not easily accessible by imaging as well as by functional electro-analytical methods up to now (e.g., Batista-García-Ramó and Fernández-Verdecia, 2018), neurobiologically inspired and computer-implemented quantitative neural models are currently an important and advantageous research tool in order to get a detailed and quantitative impression of the neurobiological functioning in speech processing (production and perception).

Finally, it needs to be stated that even if it seems to be attractive to develop computational models which probably are able to mimic the functionality of the neural system of speech production and speech perception including comprehension or of other human capabilities it should be kept in mind that all these models up to now are only of limited benefits form the viewpoint of answering fundamental research questions. And it needs to be stated that the idea of understanding speech and language disorders from a functional point of view is not new (e.g., Lichtheim, 1885). Much of the information given in chapter 4 of this paper concerning the functional specification of different types of speech and language disorders may also be deducible from box-and-arrow models (e.g., Datteri and Laudisa, 2014). But the chance to derive quantitative results like those concerning severity levels of a disorder or concerning the sensitivity level of a screening in order to detect a specific disorder or how effective a therapy method may be in strengthening a specific speaking behavior of a (model-) patient can be seized by simulating specific tasks and opens a path in the direction of increasing the efficiency of diagnosis and therapy tools. This increase in efficiency of diagnosis and therapy tools can be reached hardly by other methods as is already mentioned above in this paper (because of the need of a huge number of patients to participate in a research endeavor and because these patients then have to pass an number of screenings of therapy modules without a clear therapeutic benefit for them. Moreover, all patients recruited for this kind of studies need to be excellently diagnosed and should suffer from specific and isolated type of disorder and furthermore the exact degrees of severity of the disorder needs to be known as well, in order to get meaningful results).

The author confirms being the sole contributor of this work and has approved it for publication.