Dancheng Liu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Dancheng Liu has a total of 34 co-authors. Here are the top co-authors by number of shared publications:
- Jinjun Xiong
College of AI, Cyber and Computing, The University of Texas at San Antonio, San Antonio, TX, USA
4 shared publications - Yiyu Shi
University of Notre Dame, Notre Dame, IN, USA
4 shared publications - Ruiyang Qin
Villanova University, Villanova, USA
3 shared publications Ahmed Abbasi
University of Notre Dame, Notre Dame, USA
2 shared publications- Ahmed Abbasi
University of Notre Dame
2 shared publications - Carol A. Miller
Department of Communication Sciences and Disorders, The Pennsylvania State University, University Park, PA, USA
2 shared publications
Latest Publications
- 8 October 2026
- Published by American Speech Language Hearing Association in Language, Speech, and Hearing Services in Schools
- p. 1-13
Novel Approaches to Language Assessment Part 2: Evaluating Measures of Effort in a Sentence Recall Task for School-Age Children
Purpose: Children with developmental language disorder (DLD) produce more speech disruptions and have a slower speech rate than their peers. This study validated measures of effortful sentence production by examining their association with general language ability and differences between groups of children with DLD and typical language development (TLD). Measures of effort were expected to be related to general language ability, and children with DLD were expected to produce more stall-type disruptions (i.e., repetitions, filled pauses, and unfilled pauses) and use fewer words per minute (WPM) than their peers with TLD. Method: Audio-recorded responses to the Redmond Sentence Recall task for 139 children with TLD ( M age = 86.85 months) and 58 children with DLD ( M age = 90.64 months) were transcribed, aligned to the audio, and coded for sentence disruptions using the Codes for the Human Analysis of Transcripts form from TalkBank. A custom-made Python script computed measures of stall rate, nondisrupted rate (e.g., the percentage of utterances with no disruptions), and WPM and were examined relative to child performance on the Clinical Evaluation of Language Fundamentals–Fourth Edition. Results: When controlling for age, stall rate and nondisrupted rate were moderately correlated with general language ability, and WPM was strongly correlated with general language ability. The DLD group used significantly more stalls and fewer WPM and nondisrupted utterances than did the TLD group. Conclusions: The findings demonstrate that measures of effortful sentence production are systematically associated with language ability. Future directions for leveraging artificial intelligence to automate these measures and exploring their utility as value-added indicators of DLD are discussed.
- 8 October 2026
- Published by American Speech Language Hearing Association in Language, Speech, and Hearing Services in Schools
- p. 1-13
Novel Approaches to Language Assessment Part 2: Evaluating Measures of Effort in a Sentence Recall Task for School-Age Children
Purpose: Children with developmental language disorder (DLD) produce more speech disruptions and have a slower speech rate than their peers. This study validated measures of effortful sentence production by examining their association with general language ability and differences between groups of children with DLD and typical language development (TLD). Measures of effort were expected to be related to general language ability, and children with DLD were expected to produce more stall-type disruptions (i.e., repetitions, filled pauses, and unfilled pauses) and use fewer words per minute (WPM) than their peers with TLD. Method: Audio-recorded responses to the Redmond Sentence Recall task for 139 children with TLD ( M age = 86.85 months) and 58 children with DLD ( M age = 90.64 months) were transcribed, aligned to the audio, and coded for sentence disruptions using the Codes for the Human Analysis of Transcripts form from TalkBank. A custom-made Python script computed measures of stall rate, nondisrupted rate (e.g., the percentage of utterances with no disruptions), and WPM and were examined relative to child performance on the Clinical Evaluation of Language Fundamentals–Fourth Edition. Results: When controlling for age, stall rate and nondisrupted rate were moderately correlated with general language ability, and WPM was strongly correlated with general language ability. The DLD group used significantly more stalls and fewer WPM and nondisrupted utterances than did the TLD group. Conclusions: The findings demonstrate that measures of effortful sentence production are systematically associated with language ability. Future directions for leveraging artificial intelligence to automate these measures and exploring their utility as value-added indicators of DLD are discussed.
- 8 October 2026
- Published by American Speech Language Hearing Association in Language, Speech, and Hearing Services in Schools
- p. 1-12
Novel Approaches to Language Assessment Part 1: Artificial Intelligence–Automated Scoring of a Sentence Recall Task for School-Age Children
Purpose: Artificial intelligence (AI) technologies have the potential to enhance the delivery of speech-language pathology services. One area of practice that could benefit from AI is automated screenings for developmental language disorder (DLD) in early elementary students. This report presents the elements and processes involved in evaluating an AI-based automated scoring system using archival recordings. Method: A total of 947 audio samples of children completing the Redmond Sentence Recall (RSR) task, representing cases with and without DLD, were used. We developed an automated version of the RSR (AutoRSR), employing automated speech recognition (ASR), natural language processing, and machine learning models, to automate the tasks of transcribing recordings, tracking the number of errors children produced, scoring production accuracies based on RSR criteria, and generating pass/fail decisions based on age-referenced expectations against a reference standard (i.e., Core Language scores from the Clinical Evaluation of Language Fundamentals–Fourth Edition for a subset of 130 children). Different speech recognition models (Whisper and Reverb ASR) were applied to examine potential trade-offs. The accuracy of the AutoRSR was compared to human scoring. Results: AutoRSR with Whisper speech recognition yielded sensitivity = .768 and specificity = .784, whereas Reverb speech recognition yielded sensitivity = .873 and specificity = .730. While AutoRSR had some mismatches with human evaluators and tended to produce lower scores, it converged more closely with the reference standard than human-scored protocols. Conclusions: AI technologies have become powerful enough to automate the scoring of recordings of children's sentence recall with similar accuracy levels as human evaluators. Next steps are discussed. Supplemental Material: https://doi.org/10.23641/asha.34050876
- 8 October 2026
- Published by American Speech Language Hearing Association in Language, Speech, and Hearing Services in Schools
- p. 1-12
Novel Approaches to Language Assessment Part 1: Artificial Intelligence–Automated Scoring of a Sentence Recall Task for School-Age Children
Purpose: Artificial intelligence (AI) technologies have the potential to enhance the delivery of speech-language pathology services. One area of practice that could benefit from AI is automated screenings for developmental language disorder (DLD) in early elementary students. This report presents the elements and processes involved in evaluating an AI-based automated scoring system using archival recordings. Method: A total of 947 audio samples of children completing the Redmond Sentence Recall (RSR) task, representing cases with and without DLD, were used. We developed an automated version of the RSR (AutoRSR), employing automated speech recognition (ASR), natural language processing, and machine learning models, to automate the tasks of transcribing recordings, tracking the number of errors children produced, scoring production accuracies based on RSR criteria, and generating pass/fail decisions based on age-referenced expectations against a reference standard (i.e., Core Language scores from the Clinical Evaluation of Language Fundamentals–Fourth Edition for a subset of 130 children). Different speech recognition models (Whisper and Reverb ASR) were applied to examine potential trade-offs. The accuracy of the AutoRSR was compared to human scoring. Results: AutoRSR with Whisper speech recognition yielded sensitivity = .768 and specificity = .784, whereas Reverb speech recognition yielded sensitivity = .873 and specificity = .730. While AutoRSR had some mismatches with human evaluators and tended to produce lower scores, it converged more closely with the reference standard than human-scored protocols. Conclusions: AI technologies have become powerful enough to automate the scoring of recordings of children's sentence recall with similar accuracy levels as human evaluators. Next steps are discussed. Supplemental Material: https://doi.org/10.23641/asha.34050876
- Open Access
- 10 September 2026
- Published by Springer Nature in Nature Communications
- Vol. 17 (1)
Assessing the use of automatic speech recognition and large language models for individuals with language impairments
Large language models (LLMs) have attracted much attention for healthcare applications, demonstrating strong potential in automating conversational interactions. However, cloud-hosted systems raise privacy concerns and text-based interfaces can be difficult for children, older adults, and individuals with language impairments. Edge-deployed, voice-enabled LLM systems may address these barriers by keeping data local and using automatic speech recognition (ASR) for speech input. However, ASR transcripts often contain disfluencies, fillers, grammatical errors, and recognition noise, which may reduce downstream LLM performance for speakers with atypical language patterns. Here, we systematically evaluate edge-oriented ASR-LLM pipelines for individuals with language impairments using comparison studies and ablation experiments across aphasia, child language impairment, and dementia datasets. We identify transcript errors, repetition, noise, and input length as factors affecting system performance and deployment feasibility. These findings highlight key challenges for building more reliable and accessible speech-enabled AI systems for healthcare applications.
- 10 September 2026
- Published by Springer Nature in Nature Communications
- Vol. 17 (1)
Open AccessAssessing the use of automatic speech recognition and large language models for individuals with language impairments
Large language models (LLMs) have attracted much attention for healthcare applications, demonstrating strong potential in automating conversational interactions. However, cloud-hosted systems raise privacy concerns and text-based interfaces can be difficult for children, older adults, and individuals with language impairments. Edge-deployed, voice-enabled LLM systems may address these barriers by keeping data local and using automatic speech recognition (ASR) for speech input. However, ASR transcripts often contain disfluencies, fillers, grammatical errors, and recognition noise, which may reduce downstream LLM performance for speakers with atypical language patterns. Here, we systematically evaluate edge-oriented ASR-LLM pipelines for individuals with language impairments using comparison studies and ablation experiments across aphasia, child language impairment, and dementia datasets. We identify transcript errors, repetition, noise, and input length as factors affecting system performance and deployment feasibility. These findings highlight key challenges for building more reliable and accessible speech-enabled AI systems for healthcare applications.
- Open Access
- 12 August 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Design Automation of Electronic Systems
- Vol. 30 (5) , 1-58
Empirical Guidelines for Deploying LLMs onto Resource-constrained Edge Devices
The scaling laws have become the de facto guidelines for designing large language models (LLMs), but they were studied under the assumption of unlimited computing resources for both training and inference. As LLMs are increasingly used as personalized intelligent assistants, their customization (i.e., learning through fine-tuning) and deployment onto resource-constrained edge devices will become more and more prevalent. An urgent but open question is how a resource-constrained computing environment would affect the design choices for a personalized LLM. We study this problem empirically in this work. In particular, we consider the tradeoffs among a number of key design factors and their intertwined impacts on learning efficiency and accuracy. The factors include the learning methods for LLM customization, the amount of personalized data used for learning customization, the types and sizes of LLMs, the compression methods of LLMs, the amount of time afforded to learn, and the difficulty levels of the target use cases. Through extensive experimentation and benchmarking, we draw a number of surprisingly insightful guidelines for deploying LLMs onto resource-constrained devices. For example, an optimal choice between parameter learning and RAG may vary depending on the difficulty of the downstream task, the longer fine-tuning time does not necessarily help the model, and a compressed LLM may be a better choice than an uncompressed LLM to learn from limited personalized data.
- 12 August 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Design Automation of Electronic Systems
- Vol. 30 (5) , 1-58
Open AccessEmpirical Guidelines for Deploying LLMs onto Resource-constrained Edge Devices
The scaling laws have become the de facto guidelines for designing large language models (LLMs), but they were studied under the assumption of unlimited computing resources for both training and inference. As LLMs are increasingly used as personalized intelligent assistants, their customization (i.e., learning through fine-tuning) and deployment onto resource-constrained edge devices will become more and more prevalent. An urgent but open question is how a resource-constrained computing environment would affect the design choices for a personalized LLM. We study this problem empirically in this work. In particular, we consider the tradeoffs among a number of key design factors and their intertwined impacts on learning efficiency and accuracy. The factors include the learning methods for LLM customization, the amount of personalized data used for learning customization, the types and sizes of LLMs, the compression methods of LLMs, the amount of time afforded to learn, and the difficulty levels of the target use cases. Through extensive experimentation and benchmarking, we draw a number of surprisingly insightful guidelines for deploying LLMs onto resource-constrained devices. For example, an optimal choice between parameter learning and RAG may vary depending on the difficulty of the downstream task, the longer fine-tuning time does not necessarily help the model, and a compressed LLM may be a better choice than an uncompressed LLM to learn from limited personalized data.
- Open Access
- 24 July 2025
- Published by Springer Nature in Research Square
Analysis: Serving Individuals with Language Impairments using Automatic Speech Recognition Models and Large Language Models: Challenges and Opportunities
Large language models (LLMs) have attracted much attention for healthcare applications, demonstrating strong potential in automating conversational interactions. However, cloud-hosted LLMs pose major data privacy concerns when processing Protected Health Information. Moreover, current LLM-based systems rely on text input/output, creating substantial barriers for users, such as children and older adults, who may have difficulty typing. To mitigate these challenges, there has been growing interest in developing edge device-based, voice-enabled LLM systems. Running LLMs on edge devices minimizes the risks of PHI leaking to the cloud, while automatic speech recognition (ASR) eliminates the need for text-based inputs. Despite these advantages, existing ASRs convert speech into word-by-word text, which often contains disfluencies and fillers (e.g., ”um”, ”hum”) and grammatical errors, especially for individuals with language impairments. This noisy input can significantly degrade the performance of LLMs, yet this chained issue remains under-explored in healthcare applications. To address this critical gap, we conducted a systematic analysis through comparison studies and ablation experiments to identify key factors affecting the performance of edge-based ASR-LLM systems when used by individuals with language impairments. Furthermore, we proposed an evaluation framework for speech-enabled AI healthcare to emphasize both interpretability and robustness, paving the way for more inclusive and secure conversational healthcare solutions.
- 24 July 2025
- Published by Springer Nature in Research Square
Open AccessAnalysis: Serving Individuals with Language Impairments using Automatic Speech Recognition Models and Large Language Models: Challenges and Opportunities
Large language models (LLMs) have attracted much attention for healthcare applications, demonstrating strong potential in automating conversational interactions. However, cloud-hosted LLMs pose major data privacy concerns when processing Protected Health Information. Moreover, current LLM-based systems rely on text input/output, creating substantial barriers for users, such as children and older adults, who may have difficulty typing. To mitigate these challenges, there has been growing interest in developing edge device-based, voice-enabled LLM systems. Running LLMs on edge devices minimizes the risks of PHI leaking to the cloud, while automatic speech recognition (ASR) eliminates the need for text-based inputs. Despite these advantages, existing ASRs convert speech into word-by-word text, which often contains disfluencies and fillers (e.g., ”um”, ”hum”) and grammatical errors, especially for individuals with language impairments. This noisy input can significantly degrade the performance of LLMs, yet this chained issue remains under-explored in healthcare applications. To address this critical gap, we conducted a systematic analysis through comparison studies and ablation experiments to identify key factors affecting the performance of edge-based ASR-LLM systems when used by individuals with language impairments. Furthermore, we proposed an evaluation framework for speech-enabled AI healthcare to emphasize both interpretability and robustness, paving the way for more inclusive and secure conversational healthcare solutions.
- Open Access
- 27 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 1-9
Robust Implementation of Retrieval-Augmented Generation on Edge-based Computing-in-Memory Architectures
- 27 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 1-9
Open AccessRobust Implementation of Retrieval-Augmented Generation on Edge-based Computing-in-Memory Architectures