Automatic Speech Recognition Transcripts Research Articles

Automatic Speech Recognition in Primary Progressive Apraxia of Speech.

Transcribing disordered speech can be useful when diagnosing motor speech disorders such as primary progressive apraxia of speech (PPAOS), who have sound additions, deletions, and substitutions, or distortions and/or slow, segmented speech. Since transcribing speech can be a laborious process and requires an experienced listener, using automatic speech recognition (ASR) systems for diagnosis and treatment monitoring is appealing. This study evaluated the efficacy of a readily available ASR system (wav2vec 2.0) in transcribing speech of PPAOS patients to determine if the word error rate (WER) output by the ASR can differentiate between healthy speech and PPAOS and/or among its subtypes, whether WER correlates with AOS severity, and how the ASR's errors compare to those noted in manual transcriptions. Forty-five patients with PPAOS and 22 healthy controls were recorded repeating 13 words, 3 times each, which were transcribed manually and using wav2vec 2.0. The WER and phonetic and prosodic speech errors were compared between groups, and ASR results were compared against manual transcriptions. Mean overall WER was 0.88 for patients and 0.33 for controls. WER significantly correlated with AOS severity and accurately distinguished between patients and controls but not between AOS subtypes. The phonetic and prosodic errors from the ASR transcriptions were also unable to distinguish between subtypes, whereas errors calculated from human transcriptions were. There was poor agreement in the number of phonetic and prosodic errors between the ASR and human transcriptions. This study demonstrates that ASR can be useful in differentiating healthy from disordered speech and evaluating PPAOS severity but does not distinguish PPAOS subtypes. ASR transcriptions showed weak agreement with human transcriptions; thus, ASR may be a useful tool for the transcription of speech in PPAOS, but the research questions posed must be carefully considered within the context of its limitations. https://doi.org/10.23641/asha.26359417.

Adaptive unified defense framework for tackling adversarial audio attacks

Adversarial attacks aimed at subverting recognition systems have laid bare significant security vulnerabilities inherent in deep neural networks. In the automatic speech recognition (ASR) domain, prevailing defense mechanisms have primarily centered on pre-processing procedures to mitigate adversarial threats stemming from such attacks. However, despite their initial success, these methods have shown surprising vulnerabilities when confronted with robust and adaptive adversarial attacks. This paper proposes an adaptive unified defense framework tailored to address the challenges posed by robust audio adversarial examples. The framework comprises two pivotal components: (1) a unified pre-processing mechanism is designed to disrupt the continuity and transferability of adversarial attacks. Its objective is to thwart the consistent operation of adversarial examples across different systems or conditions, thereby enhancing the robustness of the defense. (2) an adaptive ASR transcription method is proposed to further bolster our defense strategy. Empirical experiments conducted using two benchmark audio datasets within a state-of-the-art ASR system affirm the effectiveness of our adaptive defense framework. It achieves an impressive 100% accuracy rate against representative audio attacks and consistently outperforms other state-of-the-art defense techniques, achieving an accuracy rate of 98.5% even when faced with various challenging adaptive adversarial attacks.

Analyzing the Effects of Transcription Errors on Summary Generation of Bengali Spoken Documents

Automatic speech recognition (ASR) has become an indispensable part of the AI domain, with various speech technologies reliant on it. The quality of speech recognition depends on the amount of annotated data used to train an ASR system, among other factors. For a low-resourced language, this is a severe constraint and thus ASR quality is often poor. Humans can read through text containing ASR-errors, provided the context of the sentence is preserved. Yet in cases of transcripts generated by ASR systems of low-resource languages, multiple important words are misrecognized and the context is mostly lost; discerning such a text becomes nearly impossible. This paper analyzes the types of transcription errors that occur while generating ASR transcripts of spoken documents in Bengali, an under-resourced language predominantly spoken in India and Bangladesh. The transcripts of the Bengali spoken document are generated using the ASR of Google Cloud Speech. The paper also explores if there is an effect of such transcription errors in generating speech summaries of these spoken documents. Summarization is carried out extractively; sentences are selected from the ASR-generated text of the spoken document. Speech summaries are created by aggregating the speech-segments from the original speech of the selected sentences. Subjective evaluation shows the ‘readability’ of the spoken summaries are not degraded by ASR errors, but the quality is affected due to the reliance on intermediate text-summary containing transcription errors.

Keep the noise down: On the performance of automatic speech recognition of voice-recordings in web surveys

Voice-recordings are increasingly implemented in web surveys, but the resulting audio data need to be transcribed before analysis. Since manual coding is too time- and work-intensive, researchers often rely on automatic speech recognition (ASR) systems for the transcription of the voice-recordings. However, ASR tools might create partly incorrect transcriptions and potentially change the content of responses. If the ASR performance (i.e., accuracy and validity) differs by subgroup and contextual factors, a bias is introduced in the analysis of open-ended questions. We assessed the impact of sociodemographic and contextual factors on the accuracy and validity of ASR transcriptions with data from the Longitudinal Internet Studies for the Social Sciences (LISS) panel collected in December 2020. We find that background noise reduces the accuracy and validity of ASR transcriptions. In addition, validity improved when the respondent was alone during the survey. Fortunately, we did not find any evidence of systematic differences across subgroups (age, sex, education), devices or respondent location.

Useful blunders: Can automated speech recognition errors improve downstream dementia classification?

Objectives:We aimed to investigate how errors from automatic speech recognition (ASR) systems affect dementia classification accuracy, specifically in the “Cookie Theft” picture description task. We aimed to assess whether imperfect ASR-generated transcripts could provide valuable information for distinguishing between language samples from cognitively healthy individuals and those with Alzheimer’s disease (AD). Methods:We conducted experiments using various ASR models, refining their transcripts with post-editing techniques. Both these imperfect ASR transcripts and manually transcribed ones were used as inputs for the downstream dementia classification. We conducted comprehensive error analysis to compare model performance and assess ASR-generated transcript effectiveness in dementia classification. Results:Imperfect ASR-generated transcripts surprisingly outperformed manual transcription for distinguishing between individuals with AD and those without in the “Cookie Theft” task. These ASR-based models surpassed the previous state-of-the-art approach, indicating that ASR errors may contain valuable cues related to dementia. The synergy between ASR and classification models improved overall accuracy in dementia classification. Conclusion:Imperfect ASR transcripts effectively capture linguistic anomalies linked to dementia, improving accuracy in classification tasks. This synergy between ASR and classification models underscores ASR’s potential as a valuable tool in assessing cognitive impairment and related clinical applications.

Task-Oriented Document-Grounded Dialog Systems by HLTPR@RWTH for DSTC9 and DSTC10

This paper summarizes our contributions to the document-grounded dialog tasks at the 9th and 10th Dialog System Technology Challenges (DSTC9 and DSTC10). In both iterations the task consists of three subtasks: first detect whether the current turn is knowledge seeking, second select a relevant knowledge document, and third generate a response grounded on the selected document. For DSTC9 we proposed different approaches to make the selection task more efficient. The best method, Hierarchical Selection, actually improves the results compared to the original baseline and gives a speedup of 24x. In the DSTC10 iteration of the task, the challenge was to adapt systems trained on written dialogs to perform well on noisy automatic speech recognition transcripts. Therefore, we proposed data augmentation techniques to increase the robustness of the models as well as methods to adapt the style of generated responses to fit well into the proceeding dialog. Additionally, we proposed a noisy channel model that allows for increasing the factuality of the generated responses. In addition to summarizing our previous contributions, in this work, we also report on a few small improvements and reconsider the automatic evaluation metrics for the generation task which have shown a low correlation to human judgments.

Double modals in Australian and New Zealand English

AbstractThis paper reports the first large‐scale corpus study of double modal usage in Australian and New Zealand Englishes, based on a multi‐million‐word corpus of geolocated automatic speech recognition transcripts from YouTube. Double modals are considered rare grammatical features of English, which have long been extremely difficult to observe in natural language due to low frequencies, non‐standardness, and restriction to oral speech registers. In addition, it has generally been assumed that they make up small sets of diachronically related forms, whose geographical distribution is mainly restricted to the Southern US and the North of the UK. Our results challenge these long‐standing assumptions by presenting the first inventory of double modals observed outside of these regions, thanks to computational sociolinguistic methods. Overall, we identify and map 474 double modal tokens distributed in 51 types, an unexpectedly large collection of forms used with varying frequencies across Australia and New Zealand. We consider the relevance of our results for three specific new claims concerning the diversity, complexity, and origins of double modals in English world‐wide.

AI-based approach for transcribing and classifying unstructured emergency call data: A methodological proposal.

Emergency care-sensitive conditions (ECSCs) require rapid identification and treatment and are responsible for over half of all deaths worldwide. Prehospital emergency care (PEC) can provide rapid treatment and access to definitive care for many ECSCs and can reduce mortality in several different settings. The objective of this study is to propose a method for using artificial intelligence (AI) and machine learning (ML) to transcribe audio, extract, and classify unstructured emergency call data in the Serviço de Atendimento Móvel de Urgência (SAMU) system in southern Brazil. The study used all "1-9-2" calls received in 2019 by the SAMU Novo Norte Emergency Regulation Center (ERC) call center in Maringá, in the Brazilian state of Paraná. The calls were processed through a pipeline using machine learning algorithms, including Automatic Speech Recognition (ASR) models for transcription of audio calls in Portuguese, and a Natural Language Understanding (NLU) classification model. The pipeline was trained and validated using a dataset of labeled calls, which were manually classified by medical students using LabelStudio. The results showed that the AI model was able to accurately transcribe the audio with a Word Error Rate of 42.12% using Wav2Vec 2.0 for ASR transcription of audio calls in Portuguese. Additionally, the NLU classification model had an accuracy of 73.9% in classifying the calls into different categories in a validation subset. The study found that using AI to categorize emergency calls in low- and middle-income countries is largely unexplored, and the applicability of conventional open-source ML models trained on English language datasets is unclear for non-English speaking countries. The study concludes that AI can be used to transcribe audio and extract and classify unstructured emergency call data in an emergency system in southern Brazil as an initial step towards developing a decision-making support tool.

Using HIPAA (Health Insurance Portability and Accountability Act)-Compliant Transcription Services for Virtual Psychiatric Interviews: Pilot Comparison Study.

Automatic speech recognition (ASR) technology is increasingly being used for transcription in clinical contexts. Although there are numerous transcription services using ASR, few studies have compared the word error rate (WER) between different transcription services among different diagnostic groups in a mental health setting. There has also been little research into the types of words ASR transcriptions mistakenly generate or omit. This study compared the WER of 3 ASR transcription services (Amazon Transcribe [Amazon.com, Inc], Zoom-Otter AI [Zoom Video Communications, Inc], and Whisper [OpenAI Inc]) in interviews across 2 different clinical categories (controls and participants experiencing a variety of mental health conditions). These ASR transcription services were also compared with a commercial human transcription service, Rev (Rev.Com, Inc). Words that were either included or excluded by the error in the transcripts were systematically analyzed by their Linguistic Inquiry and Word Count categories. Participants completed a 1-time research psychiatric interview, which was recorded on a secure server. Transcriptions created by the research team were used as the gold standard from which WER was calculated. The interviewees were categorized into either the control group (n=18) or the mental health condition group (n=47) using the Mini-International Neuropsychiatric Interview. The total sample included 65 participants. Brunner-Munzel tests were used for comparing independent sets, such as the diagnostic groupings, and Wilcoxon signed rank tests were used for correlated samples when comparing the total sample between different transcription services. There were significant differences between each ASR transcription service's WER (P<.001). Amazon Transcribe's output exhibited significantly lower WERs compared with the Zoom-Otter AI's and Whisper's ASR. ASR performances did not significantly differ across the 2 different clinical categories within each service (P>.05). A comparison between the human transcription service output from Rev and the best-performing ASR (Amazon Transcribe) demonstrated a significant difference (P<.001), with Rev having a slightly lower median WER (7.6%, IQR 5.4%-11.35 vs 8.9%, IQR 6.9%-11.6%). Heat maps and spider plots were used to visualize the most common errors in Linguistic Inquiry and Word Count categories, which were found to be within 3 overarching categories: Conversation, Cognition, and Function. Overall, consistent with previous literature, our results suggest that the WER between manual and automated transcription services may be narrowing as ASR services advance. These advances, coupled with decreased cost and time in receiving transcriptions, may make ASR transcriptions a more viable option within health care settings. However, more research is required to determine if errors in specific types of words impact the analysis and usability of these transcriptions, particularly for specific applications and in a variety of populations in terms of clinical diagnosis, literacy level, accent, and cultural origin.

A new corpus of geolocated ASR transcripts from Germany

AbstractThis report describes the Corpus of German Speech (CoGS), a 56-million-word corpus of automatic speech recognition transcripts from YouTube channels of local government entities in Germany. Transcripts have been annotated with latitude and longitude coordinates, making the resource potentially useful for geospatial analyses of lexical, morpho-syntactic, and pragmatic variation; this is exemplified with an exploratory geospatial analysis of grammatical variation in the encoding of past temporal reference. Additional corpus metadata include video identifiers and timestamps on individual word tokens, making it possible to search for specific discourse content or utterance sequences in the corpus and download the underlying video and audio from the web, using open-source tools. The discourse content of the transcripts in CoGS touches upon a wide range of topics, making the resource potentially interesting as a data source for research in digital humanities and social science. The report also briefly discusses the permissibility of reuse of data sourced from German municipalities for corpus-building purposes in the context of EU, German, and American law, which clearly authorize such a use case.

Feedback From Automatic Speech Recognition to Elicit Clear Speech in Healthy Speakers.

This study assessed the effectiveness of feedback generated by automatic speech recognition (ASR) for eliciting clear speech from young, healthy individuals. As a preliminary step toward exploring a novel method for eliciting clear speech in patients with dysarthria, we investigated the effects of ASR feedback in healthy controls. If successful, ASR feedback has the potential to facilitate independent, at-home clear speech practice. Twenty-three healthy control speakers (ages 23-40 years) read sentences aloud in three speaking modes: Habitual, Clear (over-enunciated), and in response to ASR feedback (ASR). In the ASR condition, we used Mozilla DeepSpeech to transcribe speech samples and provide participants with a value indicating the accuracy of the ASR's transcription. For speakers who achieved sufficiently high ASR accuracy, noise was added to their speech at a participant-specific signal-to-noise ratio to ensure that each participant had to over-enunciate to achieve high ASR accuracy. Compared to habitual speech, speech produced in the ASR and Clear conditions was clearer, as rated by speech-language pathologists, and more intelligible, per speech-language pathologist transcriptions. Speech in the Clear and ASR conditions aligned on several acoustic measures, particularly those associated with increased vowel distinctiveness and decreased speaking rate. However, ASR accuracy, intelligibility, and clarity were each correlated with different speech features, which may have implications for how people change their speech for ASR feedback. ASR successfully elicited outcomes similar to clear speech in healthy speakers. Future work should investigate its efficacy in eliciting clear speech in people with dysarthria.

Training RNN language models on uncertain ASR hypotheses in limited data scenarios

Training domain-specific automatic speech recognition (ASR) systems requires a suitable amount of data comprising the target domain. In several scenarios, such as early development stages, privacy-critical applications, or under-resourced languages, only a limited amount of in-domain speech data and an even smaller amount of manual text transcriptions, if any, are available. This motivates the study of ASR language model (LM) training on a limited amount of in-domain speech data. Early works have attempted training of n-gram LMs from ASR N-best lists and lattices but training and adaptation of recurrent neural network (RNN) LMs from ASR transcripts has not received attention. In this work, we study training and adaptation of RNN LMs using alternate, uncertain ASR hypotheses embedded in ASR confusion networks obtained from target domain speech data. We explore different methods for training the RNN LMs to deal with the uncertain input sequences. The first method extends the cross-entropy objective into a Kullback–Leibler (KL) divergence based training loss, the second method formulates a training loss based on a hidden Markov model (HMM), and the third method performs training on paths sampled from the confusion networks. These methods are applied to limited data setups including telephone and meeting conversation datasets. Performance is evaluated under two settings wherein no manual transcriptions or a small amount of manual transcriptions are available to aid the training. Moreover, a model adaptation setting is also evaluated wherein the RNN LM is pre-trained on an out-of-domain conversational corpus. Overall the sampling method for training RNN LMs on ASR confusion networks performs the best, and results in up to 12% relative reduction in perplexity on the meeting dataset as compared to training on ASR 1-best hypotheses, without any manual transcriptions. However, the perplexity reductions do not translate into equivalent WER reductions. A detailed analysis of the perplexity reductions obtained by the different methods is performed in order to understand this effect.

Multimodal attention for lip synthesis using conditional generative adversarial networks

The synthesis of lip movements is an important problem for a socially interactive agent (SIA). It is important to generate lip movements that are synchronized with speech and have realistic co-articulation. We hypothesize that combining lexical information (i.e., sequence of phonemes) and acoustic features can lead not only to models that generate the correct lip movements matching the articulatory movements, but also to trajectories that are well synchronized with the speech emphasis and emotional content. This work presents attention-based frameworks that use acoustic and lexical information to enhance the synthesis of lip movements. The lexical information is obtained from automatic speech recognition (ASR) transcriptions, broadening the range of applications of the proposed solution. We propose models based on conditional generative adversarial networks (CGAN) with self-modality attention and cross-modalities attention mechanisms. These models allow us to understand which frames are considered more in the generation of lip movements. We animate the synthesized lip movements using blendshapes. These animations are used to compare our proposed multimodal models with alternative methods, including unimodal models implemented with either text or acoustic features. We rely on subjective metrics using perceptual evaluations and an objective metric based on the LipSync model. The results show that our proposed models with attention mechanisms are preferred over the baselines on the perception of naturalness. The addition of cross-modality attentions and self-modality attentions has a significant positive impact on the performance of the generated sequences. We observe that lexical information provides valuable information even when the transcriptions are not perfect. The improved performance observed by the multimodal system confirms the complementary information provided by the speech and text modalities.

TEVL: Trilinear Encoder for Video-language Representation Learning

Pre-training model on large-scale unlabeled web videos followed by task-specific fine-tuning is a canonical approach to learning video and language representations. However, the accompanying Automatic Speech Recognition (ASR) transcripts in these videos are directly transcribed from audio, which may be inconsistent with visual information and would impair the language modeling ability of the model. Meanwhile, previous V-L models fuse visual and language modality features using single- or dual-stream architectures, which are not suitable for the current situation. Besides, traditional V-L research focuses mainly on the interaction between vision and language modalities and leaves the modeling of relationships within modalities untouched. To address these issues and maintain a small manual labor cost, we add automatically extracted dense captions as a supplementary text and propose a new trilinear video-language interaction framework TEVL (Trilinear Encoder for Video-Language representation learning). TEVL contains three unimodal encoders, a TRIlinear encOder (TRIO) block, and a temporal Transformer. TRIO is specially designed to support effective text-vision-text interaction, which encourages inter-modal cooperation while maintaining intra-modal dependencies. We pre-train TEVL on the HowTo100M and TV datasets with four task objectives. Experimental results demonstrate that TEVL can learn powerful video-text representation and achieve competitive performance on three downstream tasks, including multimodal video captioning, video Question Answering (QA), as well as video and language inference. Implementation code is available at https://github.com/Gufrannn/TEVL .

Improving Readability for Automatic Speech Recognition Transcription

Modern Automatic Speech Recognition (ASR) systems can achieve high performance in terms of recognition accuracy. However, a perfectly accurate transcript still can be challenging to read due to grammatical errors, disfluency, and other noises common in spoken communication. These readable issues introduced by speakers and ASR systems will impair the performance of downstream tasks and the understanding of human readers. In this work, we present a task called ASR post-processing for readability (APR) and formulate it as a sequence-to-sequence text generation problem. The APR task aims to transform the noisy ASR output into a readable text for humans and downstream tasks while maintaining the semantic meaning of speakers. We further study the APR task from the benchmark dataset, evaluation metrics, and baseline models: First, to address the lack of task-specific data, we propose a method to construct a dataset for the APR task by using the data collected for grammatical error correction. Second, we utilize metrics adapted or borrowed from similar tasks to evaluate model performance on the APR task. Lastly, we use several typical or adapted pre-trained models as the baseline models for the APR task. Furthermore, we fine-tune the baseline models on the constructed dataset and compare their performance with a traditional pipeline method in terms of proposed evaluation metrics. Experimental results show that all the fine-tuned baseline models perform better than the traditional pipeline method, and our adapted RoBERTa model outperforms the pipeline method by 4.95 and 6.63 BLEU points on two test sets, respectively. The human evaluation and case study further reveal the ability of the proposed model to improve the readability of ASR transcripts.

Double modals in contemporary British and Irish speech

This article reports on the use of double modals, a non-standard syntactic feature, in the contemporary speech of the UK and Ireland. Most data on the geographic extent of the feature and its combinatorial types come from surveys or acceptability ratings or from older attestations focused on northern England, Scotland or Northern Ireland, with relatively few attestations in naturalistic data and from England and Wales. Manual verification of double modals in a large corpus of geolocated Automatic Speech Recognition transcripts from YouTube videos of local government channels from the UK and Ireland shows that the feature exhibits a larger inventory of combinatorial types than has previously been found and is attested in speech from throughout the UK and Ireland. The development may be related to ongoing changes in the semantic space occupied by modal auxiliaries in English.

Using the Corpus of North American Spoken English to explore regional variation in millions of eɪ diphthongs

Increases in the quality of automatic speech recognition (ASR) systems and forced alignment algorithms have opened up new horizons for the study of acoustic properties of naturalistic speech (Coto-Solano et al., 2021). This paper describes the procedures used for the automatic extraction and alignment of audio and ASR transcript excerpts containing eɪ diphthongs from the Corpus of North American Spoken English, a 1.3-billion-word corpus of geolocated ASR transcripts of videos uploaded to YouTube channels of local government entities. In total, more than 2 million eɪ tokens were extracted from naturalistic speech in over 2000 locations in the US and Canada. The pipeline utilizes python code to extract audio segments from the DASH manifest of YouTube videos, then feeds them together with the corresponding transcript excerpt to the Montreal Forced Aligner. Targeted segments are then filtered using functions in Parselmouth, a python port of Praat. Because the underlying corpus is searchable and words are annotated with timing information, the extraction procedure is suitable for retrieving acoustic data from a range of naturalistic speech configurations (for example from discussions of particular topics or in the context of specific utterance types such as questions or replies). The corpus contains a variety of speech genres, making it suitable for studies of geophonetic and sociophonetic variation, and pipelines can be constructed to target specific phenomenona in YouTube videos, depending on the underlying research question.

DementiaBank: Theoretical Rationale, Protocol, and Illustrative Analyses.

Dementia from Alzheimer's disease (AD) is characterized primarily by a significant decline in memory abilities; however, language abilities are also commonly affected and may precede the decline of other cognitive abilities. To study the progression of language, there is a need for open-access databases that can be used to build algorithms to produce translational models sensitive enough to detect early declines in language abilities. DementiaBank is an open-access repository of transcribed video/audio data from communicative interactions from people with dementia, mild cognitive impairment (MCI), and controls. The aims of this tutorial are to (a) describe the newly established standardized DementiaBank discourse protocol, (b) describe the Delaware corpus data, and (c) provide examples of automated linguistic analyses that can be conducted with the Delaware corpus data and describe additional DementiaBank resources. The DementiaBank discourse protocol elicits four types of discourse: picture description, story narrative, procedural, and personal narrative. The Delaware corpus currently includes data from 20 neurotypical adults and 33 adults with MCI from possible AD who completed the DementiaBank discourse protocol and a cognitive-linguistic battery. Language samples were video- and audio-recorded, transcribed, coded, and uploaded to DementiaBank. The protocol materials and transcription programs can be accessed for free via the DementiaBank website. Illustrative analyses show the potential of the Delaware corpus data to help understand discourse metrics at the individual and group levels. In addition, they highlight analyses that could be used across TalkBank's other clinical banks (e.g., AphasiaBank). Information is also included on manual and automatic speech recognition transcription methods. DementiaBank is a shared online database that can facilitate research efforts to address the gaps in knowledge about language changes associated with MCI and dementia from AD. Identifying early language markers could lead to improved assessment and treatment approaches for adults at risk for dementia.

Consensus automatic speech recognition (CASR) in the California Cognitive Assessment Battery (CCAB).

Recent reports have investigated the use of automatic speech recognition (ASR) to analyze and score verbal responses in cognitive tests. ASR scoring is objective, permits the efficient computerized administration of verbal tests, and generates timestamps that enable the detailed temporal analysis of responses. However, ASR transcription accuracy varies by engine, task, and participant, and ASR can incorrectly score responses from participants with atypical speech patterns. Here we describe the speech-transcription pipeline of the California Cognitive Assessment Battery (CCAB), which incorporates consensus ASR (CASR) to produce more accurate transcripts than possible with any single ASR engine. We also developed a Transcript Review Tool (TRT) which facilitates the manual correction of mis-transcribed words in problem subjects. Figure 1 shows the CCAB speech transcription pipeline. Realtime ASR transcriptions are obtained along with the transcriptions of the digital recordings of responses using six cloud-based ASR engines (e.g., Google, etc.). Individual transcripts are then combined to produce a "consensus" transcript, and a transcription confidence measure based primarily on the agreement between ASR engines (Figure 2). If needed, "consensus" transcripts can be manually corrected using the Transcript Review Tool which enables the review of all words or just those words below a predefined CASR confidence threshold (Figure 3). ASR transcriptions were obtained from 442 healthy adults (mean age = 65.1 ±14.4) who each underwent three days of cognitive testing that included 25 verbal tests. In all, approximately 276 hours of speech were transcribed. Preliminary analyses show that CASR transcription accuracy surpassed 99% for tests with limited response sets (e.g., digit span, verbal list learning, face-name binding, etc.) and exceeded 95% for discursive speech tests (e.g., picture description and logical memory). CASR transcription is more accurate than that of any single ASR engine. When combined with the TRT, "consensus" ASR can produce error-free, timestamped transcripts that enable the detailed analysis of verbal responses from older individuals at risk of cognitive decline.

A hierarchical reasoning graph neural network for the automatic scoring of answer transcriptions in video job interviews

We address the task of automatically scoring the competency of candidates based on textual features, from the automatic speech recognition transcriptions in the asynchronous video job interviews. The key challenge is to construct the dependency relations and semantic level interaction over each question–answer (QA) pair. However, most recent studies focus on the representation of questions and answers, but ignore the dependency information and interaction between them, which is critical for QA evaluation. In this work, we propose a hierarchical reasoning graph neural network for the automatic assessment of question–answer pairs. Specifically, we construct a sentence-level relational graph neural network to capture the dependency information of sentences in or between the question and the answer. Based on these graphs, we employ a semantic-level reasoning graph attention network to model the interaction states of the current QA session. Finally, we propose a gated recurrent unit encoder to represent the temporal question–answer pairs for the final prediction. Empirical results on CHNAT (a real-world dataset) validate that our proposed model significantly outperforms matching-based benchmark models. Ablation studies and experimental results with 10 random seeds also show the effectiveness and stability of our models.