Aphasic Speech Research Articles

“The illuminating quiet”: a metaphor analysis of autobiographical descriptions of inner speech in aphasia

ABSTRACT Background Inner speech in aphasia is a rapidly expanding research area, but can be defined in numerous ways. Unlike behaviourally observable language processes such as overt word-production, inner speech is only directly available to introspection. Subjective experience thus grounds our understanding of inner speech and is a necessary starting point for investigating its definition. Descriptions of inner speech written by people with aphasia can make an important contribution to conceptual clarity in the literature. This important data source has not been systematically analysed. Our research addresses this gap. Aims This research aims to understand the subjective experience of impaired and preserved inner speech in aphasia, through analysis of the metaphors used in autobiographical accounts, and to interpret the findings in relation to the inner speech literature. Methods & Procedures This study carries out a metaphor-led discourse analysis of descriptions of inner speech in four autobiographical accounts of aphasia. Metaphorical expressions describing language processing were identified and coded, then systematic metaphors (i.e. the related concepts which are used consistently to describe a particular topic) were described. The metaphors used to describe inner speech were then analysed, with attention to patterns of use and contextual information. Results & Discussion 338 expressions made use of the main systematic metaphors which were used to describe inner speech. Two types of inner speech- Phonological Inner Speech and Dialogic Inner Speech – were described as distinct and dissociable experiences, and were described using different metaphors. Phonological Inner Speech, the internal activation of phonological representations, as used in mental repetition of a word or phrase, was described by two authors using the metaphors WORDS AS OBJECTS, MIND AS CONTAINER and INNER SPEECH AS HEARING WORDS. Dialogic Inner Speech, the use or awareness of linguistic inner reasoning or dialogue, was described by two different authors using the metaphors INNER SPEECH AS INNER VOICES/PERSONS/MONOLOGUE/DIALOGUE and APHASIA AS SILENCE/FLUID/SPIRITUAL EXPERIENCE. A double dissociation of these different definitions of inner speech is seen across two of the accounts. Different impacts on language processing and cognition were also described. Conclusions This research demonstrates that subjective accounts of inner speech can help clarify theoretical discussions and clinical implications. Although recent research within aphasiology has focused mainly on Phonological Inner Speech, we show that in these accounts the impairment of Dialogic Inner Speech was described as more salient and more explicitly as a “lack of inner voices”.

Advanced Grad-CAM Extensions for Interpretable Aphasia Speech Keyword Classification: Bridging the Gap in Impaired Speech with XAI

Advanced Grad-CAM Extensions for Interpretable Aphasia Speech Keyword Classification: Bridging the Gap in Impaired Speech with XAI

The Flu-ID: A New Evidence-Based Method of Assessing Fluency in Aphasia.

Assessing fluency in aphasia is diagnostically important for determining aphasia type and severity and therapeutically important for determining appropriate treatment targets. However, wide variability in the measures and criteria used to assess fluency, as revealed by a recent survey of clinicians (Gordon & Clough, 2022), results in poor reliability. Furthermore, poor specificity in many fluency measures makes it difficult to identify the underlying impairments. Here, we introduce the Flu-ID Aphasia, an evidence-based tool that provides a more informative method of assessing fluency by capturing the range of behaviors that can affect the flow of speech in aphasia. The development of the Flu-ID was based on prior evidence about factors underlying fluency (Clough & Gordon, 2020; Gordon & Clough, 2020) and clinical perceptions about the measurement of fluency (Gordon & Clough, 2022). Clinical utility is maximized by automated counting of fluency behaviors in an Excel template. Reliability is maximized by outlining thorough guidelines for transcription and coding. Eighteen narrative samples representing a range of fluency were coded independently by the authors to examine the Flu-ID's utility, reliability, and validity. Overall reliability was very good, with point-to-point agreement of 86% between coders. Ten of the 12 dimensions showed good to excellent reliability. Validity analyses indicated that Flu-ID scores were similar to clinician ratings on some dimensions, but differed on others. Possible reasons and implications of the discrepancies are discussed, along with opportunities for improvement. The Flu-ID assesses fluency in aphasia using a consistent and comprehensive set of measures and semi-automated procedures to generate individual fluency profiles. The profiles generated in the current study illustrate how similar ratings of fluency can arise from different underlying impairments. Supplemental materials include an analysis template, extensive guidelines for transcription and coding, a completed sample, and a quick reference guide. https://doi.org/10.23641/asha.27078199.

Interdisciplinary approaches to understanding the inner speech, with emphasis on the role of incorporating clinical data.

Neuroscience has largely conceptualized inner speech, sometimes called covert speech, as being a part of the language system, namely, a precursor to overt speech and/or speech without the motor component (impoverished motor speech). Yet interdisciplinary work has strongly suggested that inner speech is multidimensional and situated within the language system as well as in more domain general systems. By leveraging evidence from philosophy, linguistics, neuroscience and cognitive science, we argue that neuroscience can gain a more comprehensive understanding of inner speech processes. We will summarize the existing knowledge on the traditional approach to understanding the neuroscience of inner speech, which is squarely through the language system, before discussing interdisciplinary approaches to understanding the cognitive, linguistic and neural substrates/mechanisms that may be involved in inner speech. Given our own interests in inner speech after brain injury, we finish by discussing the theoretical and clinical benefits of researching inner speech in aphasia through an interdisciplinary lens.

Deictic vs. anaphoric pronouns: a comparison of fluent and non-fluent aphasia in English and Tagalog

ABSTRACT Deictic pronouns refer directly to the extralinguistic world (as in Stop her!), while anaphoric pronouns refer to elements introduced in discourse (as in My best friend is Mary. John adores her.). The distinction is central to linguistic theory and may be central also for understanding aphasia. This paper compares the production of deictic and anaphoric pronouns in fluent aphasia and non-fluent aphasia. It does so both for English and for Tagalog, which allows extensive omission of pronominal arguments. For both languages, the results show a dissociation between deictic and anaphoric pronouns. Non-fluent aphasic speech displays a lower proportion of anaphoric pronouns to non-pronouns than non-brain-damaged speech, while the proportion of deictic pronouns to non-pronouns is either not different (Tagalog) or higher (English). In contrast, fluent aphasic speech displays a higher proportion of deictic pronouns to non-pronouns, while the proportion of anaphoric pronouns is either not different (Tagalog) or lower (English).

LLM-Based Text Prediction and Question Answer Models for Aphasia Speech

LLM-Based Text Prediction and Question Answer Models for Aphasia Speech

Multilingual Aphasia Speech Analysis with Machine Learning

Aphasia is an acquired language disorder that occurs after brain injury such as stroke, head trauma or tumor. People with aphasia (PWA) may have trouble speaking or under-standing speech. If diagnosed early, aphasia is often treatable, and communication can be improved with speech therapy. Early detection and evaluation of aphasia is crucial for the treatment and recovery. This paper reports a preliminary study of multilingual aphasia speech evaluation. In this study, the characteristics of speech from PWA and healthy controls are compared from both acoustic and linguistic perspectives. Multiple acoustic features are extracted from aphasic and normal speech to build a language independent aphasic speech detection model. The model achieved good aphasic speech detection performance on both English and Mandarin test sets.

A Comparative Investigation of Automatic Speech Recognition Platforms for Aphasia Assessment Batteries.

The rehabilitation of aphasics is fundamentally based on the assessment of speech impairment. Developing methods for assessing speech impairment automatically is important due to the growing number of stroke cases each year. Traditionally, aphasia is assessed manually using one of the well-known assessment batteries, such as the Western Aphasia Battery (WAB), the Chinese Rehabilitation Research Center Aphasia Examination (CRRCAE), and the Boston Diagnostic Aphasia Examination (BDAE). In aphasia testing, a speech-language pathologist (SLP) administers multiple subtests to assess people with aphasia (PWA). The traditional assessment is a resource-intensive process that requires the presence of an SLP. Thus, automating the assessment of aphasia is essential. This paper evaluated and compared custom machine learning (ML) speech recognition algorithms against off-the-shelf platforms using healthy and aphasic speech datasets on the naming and repetition subtests of the aphasia battery. Convolutional neural networks (CNN) and linear discriminant analysis (LDA) are the customized ML algorithms, while Microsoft Azure and Google speech recognition are off-the-shelf platforms. The results of this study demonstrated that CNN-based speech recognition algorithms outperform LDA and off-the-shelf platforms. The ResNet-50 architecture of CNN yielded an accuracy of 99.64 ± 0.26% on the healthy dataset. Even though Microsoft Azure was not trained on the same healthy dataset, it still generated comparable results to the LDA and superior results to Google's speech recognition platform.

An open dataset of connected speech in aphasia with consensus ratings of auditory-perceptual features.

Auditory-perceptual rating of connected speech in aphasia (APROCSA) involves trained listeners rating a large number of perceptual features of speech samples, and has shown promise as an approach for quantifying expressive speech and language function in individuals with aphasia. The aim of this study was to obtain consensus ratings for a diverse set of speech samples, which can then be used as training materials for learning the APROCSA system. Connected speech samples were recorded from six individuals with chronic post-stroke aphasia. A segment containing the first five minutes of participant speech was excerpted from each sample, and 27 features were rated on a five-point scale by five researchers. The researchers then discussed each feature in turn to obtain consensus ratings. Six connected speech samples are made freely available for research, education, and clinical uses. Consensus ratings are reported for each of the 27 features, for each speech sample. Discrepancies between raters were resolved through discussion, yielding consensus ratings that can be expected to be more accurate than mean ratings. The dataset will provide a useful resource for scientists, students, and clinicians to learn how to evaluate aphasic speech samples with an auditory-perceptual approach.

Специфика речевой межполушарной асимметрии при афазических расстройствах и ее влияние на количественные показатели речевой перцепции и динамику афазии

Introduction. In neuropsychological practice the dichotic listening task has been widely used for more than 50 years as a valid test for determining the laterality of auditory speech asymmetry. Theoretical analysis. Most of the studies of the dichotic listening task in aphasia are characterized by small sample groups, a wide demographic features of them, lack attention to the type and aphasia severity, a size of local brain damage as well. Without taking into account these important factors it is difficult to answer to some questions: 1. whether the vector of auditory speech asymmetry is determined by the age, type and aphasia severity; 2. whether the laterality of auditory speech asymmetry affects the dynamics of aphasia recovery. The aim of the study: the connection between quantitative parameters of the dichotic listening task and the dynamics of aphasia regression. Empirical analysis. Material. The study involved patients (N = 109) with traumatic and post-stroke efferent motor (50.7 ± 10.1 years) and acoustic mnestic (54.4 ± 8.2 years) aphasia of mild and moderate severity of speech disorders. Methods. Kimura’s dichotic listening task modified by B. S. Kotik; the quantitative assessment of speech in aphasia L. S. Tsvetkova, T. V. Akhutina, N. M. Pylaeva. The speech assessment was carried out twice: before and after a 6-week course of speech rehabilitation. Results. Between efferent motor and acoustic mnestic aphasia groups the difference in the laterality of ear superiority was revealed. The efferent motor aphasia group shows the ambidextrous pattern in auditory perception. The dichotic listening task efficiency increases with a burst of the left ear productivity and vice versa with the right ear. The acoustic mnestic aphasia group presents a pronounced effect of the right ear suppression. Its productivity is significantly lower than the productivity of the left ear and the same indicator of the efferent motor aphasia group. The effectiveness of the dichotic listening task in the acoustic mnestic aphasia group does not exhibit a correlation with the parameters of productivity of the right or left ear. The aphasia groups vary in the dynamics of aphasia regression. In the acoustic mnestic aphasia group, this dynamics was 2.5 times higher than in the efferent motor aphasia group. Conclusion. The right hemisphere laterality in auditory speech perception correlates positively with the dynamics of aphasia regression. Further research should find an answer to the question of the constellations of significant factors and their hierarchical relationships that could determine the positive dynamics of recovery of aphasic disorders.

Automatic Assessment of Aphasic Speech Sensed by Audio Sensors for Classification into Aphasia Severity Levels to Recommend Speech Therapies.

Aphasia is a type of speech disorder that can cause speech defects in a person. Identifying the severity level of the aphasia patient is critical for the rehabilitation process. In this research, we identify ten aphasia severity levels motivated by specific speech therapies based on the presence or absence of identified characteristics in aphasic speech in order to give more specific treatment to the patient. In the aphasia severity level classification process, we experiment on different speech feature extraction techniques, lengths of input audio samples, and machine learning classifiers toward classification performance. Aphasic speech is required to be sensed by an audio sensor and then recorded and divided into audio frames and passed through an audio feature extractor before feeding into the machine learning classifier. According to the results, the mel frequency cepstral coefficient (MFCC) is the most suitable audio feature extraction method for the aphasic speech level classification process, as it outperformed the classification performance of all mel-spectrogram, chroma, and zero crossing rates by a large margin. Furthermore, the classification performance is higher when 20 s audio samples are used compared with 10 s chunks, even though the performance gap is narrow. Finally, the deep neural network approach resulted in the best classification performance, which was slightly better than both K-nearest neighbor (KNN) and random forest classifiers, and it was significantly better than decision tree algorithms. Therefore, the study shows that aphasia level classification can be completed with accuracy, precision, recall, and F1-score values of 0.99 using MFCC for 20 s audio samples using the deep neural network approach in order to recommend corresponding speech therapy for the identified level. A web application was developed for English-speaking aphasia patients to self-diagnose the severity level and engage in speech therapies.

On Fluency: Perspectives From Aphasiology and Second Language Acquisition Research

Fluency is a long-debated notion in linguistics, as it plays a crucial role in second language acquisition (SLA) research and applied linguistics. Starting from some recent discussions on the topic, this paper examines whether aphasiology could benefit from insights from these fields of study. First, we will summarize the most relevant views on fluency concerning aphasic speech. As we will see, every account of fluency has two common features: some reference to speech timing, and some reference to anomia. Nonetheless, fluency as a whole still remains nothing more than an intuition, since no model has implemented a satisfactory definition of it. We argue that the reason behind this decades-long uncertainty is because its two main components -time and retrieval- pertain to two distinct levels -namely speech and language- and therefore cannot be judged on the same scale. To show our point, we will briefly pass in review studies on time patterns in aphasic speech and see whether it is possible to draw a distinctive behavior compared to non-pathological speech. We will then move to anomia and see how the matter has been approached in second language research, especially in fluency modeling. While certainly not relatable on several issues, we believe that some key outcomes in linguistic fluency can be adapted to aphasia research, and offer new perspectives on aphasic speech evaluation and understanding.

BONEs not CATs attract DOGs: Semantic context effects for picture naming in the lesioned language network

The breakdown of rapid and accurate retrieval of words is a hallmark of aphasic speech and a prime target of therapeutic intervention. Complementary, psycho- and neurolinguistic research have developed a spectrum of models, how and by which neuronal network uncompromised speakers can rely on remarkable lexical retrieval capacities. Motivated by both lines of research we invited 32 participants with a chronic left hemispheric brain lesion to name pictures in the presence of distractor words. This picture-word-interference (PWI) paradigm is widely used in psycho- and neurolinguistic research.We find that also after brain lesion categorically related words (CAT → [dog]picture) impede naming, while associatively related words (BONE → [dog]picture) ease access, when compared to unrelated distractor words. The effects largely affecting latencies in neurotypical populations, are reproduced for error rate in our participants with lesions in the language network. Unsurprisingly, overall naming abilities varied greatly across patients. Notably, however, the two effects (categorical interference / associative facilitation) differ between participants. Correlating performance with lesion patterns we find support for the notion of a divergence of brain areas affording different aspects of the task: (i) lesions in the left middle temporal gyurs (MTG) deteriorate overall naming, confirming previous work; more notably, (ii) lesions comprising the inferior frontal hub (inferior frontal gyrus, IFG) of the language-network increase the interference effect for the categorical condition; on the contrary, (iii) lesions to the mid-to-posterior temporal hub (posterior middle and superior temporal gyri, pMTG/ pSTG) increase the facilitatory effect for the associative condition on error rates.The findings can be accommodated in a neuro-linguistic framework, which localizes lexical activation but also lexical interference in posterior parts of the language network (pMTG/pITG); conversely, selection between co-activated categorically related entries is afforded by frontal language areas (IFG). While purely experimental in nature our study highlights that lesion site differentially influences specific aspects of word retrieval. Since confrontational naming is a cornerstone of aphasia rehabilitation, this may be of note when designing and evaluating novel therapeutic regimes.

Improving Aphasic Speech Recognition by Using Novel Semi-Supervised Learning Methods on AphasiaBank for English and Spanish

Automatic speech recognition in patients with aphasia is a challenging task for which studies have been published in a few languages. Reasonably, the systems reported in the literature within this field show significantly lower performance than those focused on transcribing non-pathological clean speech. It is mainly due to the difficulty of recognizing a more unintelligible voice, as well as due to the scarcity of annotated aphasic data. This work is mainly focused on applying novel semi-supervised learning methods to the AphasiaBank dataset in order to deal with these two major issues, reporting improvements for the English language and providing the first benchmark for the Spanish language for which less than one hour of transcribed aphasic speech was used for training. In addition, the influence of reinforcing the training and decoding processes with out-of-domain acoustic and text data is described by using different strategies and configurations to fine-tune the hyperparameters and the final recognition systems. The interesting results obtained encourage extending this technological approach to other languages and scenarios where the scarcity of annotated data to train recognition models is a challenging reality.

Neural bases of elements of syntax during speech production in patients with aphasia

The ability to string together words into a structured arrangement capable of conveying nuanced information is key to speech production. The assessment of the neural bases for structuring sentences has been challenged by the need of experts to delineate the aberrant morphosyntactic structures in aphasic speech. Most studies have relied on focused tasks with limited ecological validity. We characterized syntactic complexity during connected speech produced by patients with chronic post-stroke aphasia. We automated this process by employing Natural Language Processing (NLP). We conducted voxel-based and connectome-based lesion-symptom mapping to identify brain regions crucially associated with sentence production and syntactic complexity. Posterior-inferior aspects of left frontal and parietal lobes, as well as white matter tracts connecting these areas, were essential for syntactic complexity, particularly the posterior inferior frontal gyrus. These findings suggest that sentence structuring during word production depends on the integrity of Broca’s area and the dorsal stream of language processing.

Simulating phonological and semantic impairment of English tense inflection with linear discriminative learning

AbstractThis study applies the computational theory of the ‘discriminative lexicon’ (Baayen, Chuang, and Blevins, 2019) to the modeling of the production of English verbs in aphasic speech. Under semantic impairment, speakers have been reported to have greater difficulties with irregular verbs, whereas speakers with phonological impairment are described as having greater problems with regulars.Joanisse and Seidenberg (1999)were able to model this dissociation, but only by adding noise to the semantic units of their model. We report two simulation studies in which topographically coherent regions of phonological and semantic networks were selectively damaged. Our model replicated the main findings, including the high variability in the consequences of brain lesions for speech production. Importantly, our model generated these results without having to lesion the semantic system more than the phonological system. The model’s success hinges on the use of a corpus-based distributional vector space for representing verbs’ meanings. Irregular verbs have denser semantic neighborhoods than do regular verbs (Baayen and Moscoso del Prado Martín, 2005). In our model this renders irregular verbs more fragile under semantic impairment. These results provide further support for the central idea underlying the discriminative lexicon: that behavioral patterns can, to a considerable extent, be understood as emerging from the distributional properties of a language and basic principles of human learning.

Factor Analysis of Spontaneous Speech in Aphasia.

Purpose Spontaneous speech tasks are critically important for characterizing spoken language production deficits in aphasia and for assessing the impact of therapy. The utility of such tasks arises from the complex interaction of linguistic demands (word retrieval, sentence formulation, articulation). However, this complexity also makes spontaneous speech hugely variable and difficult to assess. The current study aimed to simplify the problem by identifying latent factors underlying performance in spontaneous speech in aphasia. The ecological validity of the factors was examined by examining how well the factor structures corresponded to traditionally defined aphasia subtypes. Method A factor analysis was conducted on 17 microlinguistic measures of narratives from 274 individuals with aphasia in AphasiaBank. The resulting factor scores were compared across aphasia subtypes. Supervised (linear discriminant analysis) and unsupervised (latent profile analysis) classification techniques were then conducted on the factor scores and the solutions compared to traditional aphasia subtypes. Results Six factors were identified. Two reflected aspects of fluency, one at the phrase level (Phrase Building) and one at the narrative level (Narrative Productivity). Two other factors reflected the accuracy of productions, one at the word level (Semantic Anomaly) and one at the utterance level (Grammatical Error). The other two factors reflected the complexity of sentence structures (Grammatical Complexity) and the use of repair behaviors (Repair), respectively. Linear discriminant analyses showed that only about two thirds of speakers were classified correctly and that misclassifications were similar to disagreements between clinical diagnoses. The most accurately diagnosed syndromes were the largest groups-Broca's and anomic aphasia. The latent profile analysis also generated profiles similar to Broca's and anomic aphasia but separated some subtypes according to severity. Conclusions The factor solution and the classification analyses reflected broad patterns of spontaneous speech performance in a large and representative sample of individuals with aphasia. However, such data-driven approaches present a simplified picture of aphasia patterns, much as traditional syndrome categories do. To ensure ecological validity, a hybrid approach is recommended, balancing population-level analyses with examination of performance at the level of theoretically specified subgroups or individuals. Supplemental Material https://doi.org/10.23641/asha.13232354.

An Efficient Deep Learning Based Method for Speech Assessment of Mandarin-Speaking Aphasic Patients.

Speech assessment is an important part of the rehabilitation process for patients with aphasia (PWA). Mandarin speech lucidity features such as articulation, fluency, and tone influence the meaning of the spoken utterance and overall speech clarity. Automatic assessment of these features is important for an efficient assessment of the aphasic speech. Hence, in this paper, a standardized automatic speech lucidity assessment method for Mandarin-speaking aphasic patients using a machine learning based technique is presented. The proposed assessment method adopts the Chinese Rehabilitation Research Center Aphasia Examination (CRRCAE) standard as a guideline. Quadrature based high-resolution time-frequency images with a convolutional neural network (CNN) are utilized to develop a method that can map the relationship between the severity level of aphasic patients' speech and the three speech lucidity features. The results show a linear relationship with statistically significant correlations between the normalized true-class output activations (TCOA) of the CNN model and patients' articulation, fluency, and tone scores, i.e., 0.71 (p < 0.001), 0.60 (p < 0.001) and 0.58 (p < 0.001), respectively. The linearity of the proposed Mandarin aphasic speech assessment method and its significant correlation with the speech severity levels show the efficacy of the method in predicting the severity of impaired Mandarin speech. The outcome of this research envisages assisting speech-language pathologists in Mandarin-speech impairment assessment and promoting early support discharge; hence could alleviate the stress that the healthcare system is currently experiencing in China nationwide. The framework of the proposed Mandarin aphasic speech assessment method can be readily extended to other languages.

Impact of Combined Transcranial Direct Current Stimulation and Speech-language Therapy on Spontaneous Speech in Aphasia: A Randomized Controlled Double-blind Study.

Aphasia recovery depends on neural reorganization, which can be enhanced by speech-language therapy and noninvasive brain stimulation. Several studies suggested that transcranial direct current stimulation (tDCS) associated with speech-language therapy may improve verbal performance evaluated by analytic tests, but none focused on spontaneous speech. We explored the effect of bihemispheric tDCS on spontaneous speech in patients with poststroke aphasia. In this multicentric controlled randomized cross-over double-blind study, we included 10 patients with poststroke aphasia (4 had aphasia >6months and 6 with aphasia <6months). We combined the sessions of speech-language therapy and bihemispheric tDCS (2 mA, 20min). After three baseline speech evaluations (1/week), two different conditions were randomly consecutively proposed: active and sham tDCS over 3weeks with 1week of washout in between. The main outcome measure was the number of different nouns used in 2min to answer the question "what is your job." There was no significant difference between conditions concerning the main outcome measure (p = .47) nor in the number of verbs, adjectives, adverbs, pronouns, repetitions, blank ideas, ideas, utterances with grammatical errors or paraphasias used. Other cognitive functions (verbal working memory, neglect, or verbal fluency) were not significantly improved in the tDCS group. No adverse events occurred. Our results differed from previous studies using tDCS to improve naming in patients with poststroke aphasia possibly due to bihemispheric stimulation, rarely used previously. The duration of the rehabilitation period was short given the linguistic complexity of the measure. This negative result should be confirmed by larger studies with ecological measures.

Automatic Assessment of Speech Impairment in Cantonese-speaking People with Aphasia.

Aphasia is a common type of acquired language impairment resulting from dysfunction in specific brain regions. Analysis of narrative spontaneous speech, e.g., story-telling, is an essential component of standardized clinical assessment on people with aphasia (PWA). Subjective assessment by trained speech-language pathologists (SLP) have many limitations in efficiency, effectiveness and practicality. This paper describes a fully automated system for speech assessment of Cantonese-speaking PWA. A deep neural network (DNN) based automatic speech recognition (ASR) system is developed for aphasic speech by multi-task training with both in-domain and out-of-domain speech data. Story-level embedding and Siamese network are applied to derive robust text features, which can be used to quantify the difference between aphasic speech and unimpaired one. The proposed text features are combined with conventional acoustic features to cover different aspects of speech and language impairment in PWA. Experimental results show a high correlation between predicted scores and subject assessment scores. The best correlation value achieved with ASR-generated transcription is .827, as compared with .844 achieved with manual transcription. The Siamese network significantly outperforms story-level embedding in generating text features for automatic assessment.