Paper presented at the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (2026)

We address catastrophic forgetting (CF) in automatic speaker verification (SV) during domain adaptation from Adult English (VoxCeleb) to Child English (MyST) and Adult Chinese (CNCeleb), under the assumption of no access to VoxCeleb data. We leverage embeddings of target-data inputs extracted from the pretrained VoxCeleb model, hypothesizing that these embeddings retain source-domain knowledge without requiring source data. These target-data embeddings extracted from the pretrained VoxCeleb model are referred to as pretrained-model embeddings (PMEs). We propose a cross-attentive (CA) adapter that reduces CF by dynamically balancing information between learnable target-domain embeddings and PMEs during adaptation. We design two regularization strategies: (i) K-means-based elastic weight consolidation (K-EWC), where clustered PMEs provide pseudo labels for parameter-importance estimation, and (ii) a moment-matching (MM) loss that constrains learnable target-domain embeddings to remain close to the PME distribution. We evaluate across target data training splits of increasing duration. Results show that our approaches consistently reduce CF by not degrading performance (%EER) on the source domain regardless of target data size. [This paper was published in: IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) Proceedings, 2026, pp. 18877-18881.]

Descriptors: Memory, Speech Communication, English, Training Objectives, Natural Language Processing, Models, Training, Artificial Intelligence

View on ERIC

Paper presented at the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (2026)

Self-supervised learning (SSL) models have achieved impressive results across many speech tasks, yet child automatic speech recognition (ASR) remains challenging due to limited data and pretraining domain mismatch. Fine-tuning SSL models on child speech induces shifts in the representation space. We hypothesize that delta SSL embeddings, the differences between embeddings from a fine-tuned model and those from its pre-trained counterpart, encode task-specific information that complements fine-tuned features from another SSL model. We evaluate multiple fusion strategies on the MyST children’s corpus with different models. Results show that delta embedding fusion with WavLM yields up to a 10% relative WER reduction for HuBERT and a 4.4% reduction for W2V2, compared to fine-tuned embedding fusion. Notably, fusing WavLM with delta W2V2 embeddings achieves a WER of 9.64, setting a new state-of-the-art among SSL models on the MyST corpus. These findings demonstrate the effectiveness of delta embeddings and highlight feature fusion as a promising direction for advancing child ASR. [This paper was published in: IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) Proceedings, 2026, pp. 18417-18421.]

Descriptors: Independent Study, Speech Communication, Speech Evaluation, Word Recognition, Children

View on ERIC

Paper presented at the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (2026)

Neural audio codecs are widely used for audio compression and can be integrated into token-based language models. Traditional codecs preserve acoustic details well but lack semantic information. Recent hybrid codecs attempt to incorporate semantic information through distillation, but this often degrades reconstruction performance, making it difficult to achieve both. To address this limitation, we introduce STACodec, a unified codec that integrates semantic information from self-supervised learning (SSL) models into the first layer of residual vector quantization (RVQ-1) via semantic token assignment (STA). To further eliminate reliance on SSL-based semantic tokenizers and improve efficiency during inference, we propose a semantic pre-distillation (SPD) module, which predicts semantic tokens directly for assignment to the first RVQ layer during inference. Experimental results show that STACodec outperforms existing hybrid codecs in both audio reconstruction and downstream semantic tasks, demonstrating a better balance between acoustic fidelity and semantic capability. [This paper was published in: IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) Proceedings, 2026, pp. 17032-17036.]

Descriptors: Semantics, Assignments, Acoustics, Fidelity, Audio Equipment, Computer Software, Algorithms, Information Storage, Natural Language Processing

View on ERIC

Paper presented at Interspeech 2025 (Rotterdam, The Netherlands, Aug 17-21, 2025)

Automatic Speech Recognition (ASR) systems struggle with child speech due to its distinct acoustic and linguistic variability and limited availability of child speech datasets, leading to high transcription error rates. While ASR error correction (AEC) methods have improved adult speech transcription, their effectiveness on child speech remains largely unexplored. To address this, we introduce CHSER, a Generative Speech Error Correction (GenSEC) dataset for child speech, comprising 200K hypothesis-transcription pairs spanning diverse age groups and speaking styles. Results demonstrate that fine-tuning on the CHSER dataset achieves up to a 28.5% relative WER reduction in a zero-shot setting and a 13.3% reduction when applied to fine-tuned ASR systems. Additionally, our error analysis reveals that while GenSEC improves substitution and deletion errors, it struggles with insertions and child specific disfluencies. These findings highlight the potential of GenSEC for improving child ASR. [This paper was published in: Interspeech 2025 Proceedings, 2025, pp. 2895-2899.]

Descriptors: Child Language, Artificial Intelligence, Man Machine Systems, Natural Language Processing, Error Analysis (Language), Error Correction, Program Effectiveness, Automation

View on ERIC

Paper presented at the Annual Workshop on Speech and Language Technology in Education (SLaTE) (10th, Nijmegen, The Netherlands, Aug 22-24, 2025)

This paper explores the use of automatic speech recognition (ASR) and large language models (LLMs) for automated scoring and feedback generation in spoken language assessment. We design a three stage pipeline that (1) optimizes ASR hypotheses from student speech, (2) performs task-based scoring using LLMs, and (3) generates natural language feedback justifying each score. We evaluate this pipeline using audio responses from 3rd-8th grade students in the Atlanta, Georgia area, recorded as part of the Test of Narrative Language. Our results show that LLMs can reliably replicate expert annotations while providing interpretable feedback. We further analyze model performance across demographic factors, including dialect and reading proficiency, to assess equity. Our findings demonstrate the promise of ASR and LLMs for robust, explainable, and fair assessment of children’s spoken narratives. [This paper was published in: 10th Workshop on Speech and Language Technology in Education (SLaTE) Proceedings, International Speech Communication Association (ISCA), 2025, pp. 1-5.]

Descriptors: Artificial Intelligence, Automation, Natural Language Processing, Language Tests, Speech Communication, Elementary School Students, Secondary School Students, Feedback (Response), Computer Assisted Testing, Evaluation Methods, Dialects

View on ERIC

Paper presented at the Workshop on Child Computer Interaction — WOCCI 2025 (Nijmegen, The Netherlands, Aug 22-24, 2025)

In speaker verification (SV), the acoustic mismatch between children’s and adults’ speech leads to suboptimal performance when adult-trained SV systems are applied to children’s speaker verification (C-SV). While domain adaptation techniques can enhance performance on C-SV tasks, they often do so at the expense of significant degradation in performance on adults’ SV (A-SV) tasks. In this study, we propose an Age Agnostic Speaker Verification (AASV) system that achieves robust performance across both C-SV and A-SV tasks. Our approach employs a domain classifier to disentangle age-related attributes from speech and subsequently expands the embedding space using the extracted domain information, forming a unified speaker representation that is robust and highly discriminative across age groups. Experiments on the OGI and VoxCeleb datasets demonstrate the effectiveness of our approach in bridging SV performance disparities, laying the foundation for inclusive and age-adaptive SV systems.

Descriptors: Speech Communication, Acoustics, Children, Speech Evaluation, Age Groups, Adults, Classification

View on ERIC

Paper presented at the Workshop on Child Computer Interaction — WOCCI 2025 (Nijmegen, The Netherlands, Aug 22-24, 2025)

Speaker Verification (SV) systems trained on adults speech often underperform on children’s SV due to the acoustic mismatch, and limited children speech data makes fine-tuning not very effective. In this paper, we propose an innovative framework, a Gated Linear Unit adapter with Iterative Fine-Tuning (G-IFT), to enhance knowledge transfer efficiency between the high-resource adults speech domain and the low-resource children’s speech domain. In this framework, a Gated Linear Unit adapter is first inserted between the pre-trained speaker embedding model and the classifier. Then the classifier, adapter, and pre-trained speaker embedding model are optimized sequentially in an iterative way. This framework is agnostic to the type of the underlying architecture of the SV system. Our experiments on ECAPA-TDNN, ResNet, and X-vector architectures using the OGI and MyST datasets demonstrate that the G-IFT framework yields consistent reductions in Equal Error Rates compared to baseline methods.

Descriptors: Children, Speech Communication, Speech Evaluation, Models, Classification, Acoustics

View on ERIC

International Educational Data Mining Society

, Paper presented at the International Conference on Educational Data Mining (EDM) (18th, Palermo, Italy, Jul 20-23, 2025)

High-quality feedback requires understanding of a student’s work, insights into what concepts would help them improve, and language that matches the preferences of the specific teaching team. While Large Language Models (LLMs) can generate coherent feedback, adapting these responses to align with specific teacher preferences remains an open challenge. We present a method for aligning LLM-generated feedback with teacher preferences using Direct Preference Optimization (DPO). We integrate preference data collection into the grading process. This creates a self-improving pipeline keeping the teacher-in-the-loop to ensure feedback quality and maintain teacher autonomy. To evaluate effectiveness, we conducted a blind controlled study where expert evaluators compared feedback from multiple models on anonymized student submissions. Evaluators consistently preferred feedback from our DPO model over GPT-4o. We deployed the system in two offerings of a large university course, with nearly 300 students and over 10 teaching assistants per term, demonstrating its feasibility in real classroom settings. We share strategies for automated performance monitoring using critic models. We explore methods for examining fairness across protected demographics. [For the complete proceedings, see ED675583.]

Descriptors: Feedback (Response), Artificial Intelligence, Teacher Attitudes, Preferences, Data Collection, Grading, Models, Evaluation, Automation, Progress Monitoring, Higher Education

View on ERIC

Computers and Education: Artificial Intelligence v9 Article 100484 2025

Dialogic reading, a technique in which adults and children engage in interactive discussions around a story, has been shown to improve children’s language and literacy development. Despite its evidence-based benefits, its adoption among families with English as a Foreign Language (EFL) backgrounds has been particularly challenging due to limited English proficiency, restricted conversational skills, and a low inclination to read in English. This paper presents “Storio”, an e-book integrated with a bilingual large language model (LLM)-based conversational agent named “Mia”, used as a design probe to investigate interactions between EFL children (N=17) and the agent, and to gather insights from parents (N=19) and educators (N=2). The findings indicate that the bilingual agent effectively supports language output, fosters interactive experiences, and promotes language skills. The study offers valuable design implications for the development of LLM-based and children’s interactive e-books tailored to the needs of children with diverse linguistic and cultural backgrounds.

Descriptors: Artificial Intelligence, Technology Uses in Education, English (Second Language), Electronic Books, Bilingualism, Reading, Parents, Young Children, Chinese, Teachers

View on ERIC

Paper presented at the Workshop on Child Computer Interaction — WOCCI 2025 (Nijmegen, The Netherlands, Aug 22-24, 2025)

Despite advancements in ASR, child speech recognition remains challenging due to acoustic variability and limited annotated data. While fine-tuning adult ASR models on child speech is common, comparisons with flat-start training remain underexplored. We compare flat-start training across multiple datasets, SSL representations (WavLM, XEUS), and decoder architectures. Our results show that SSL representations are biased toward adult speech, with flat-start training on child speech mitigating these biases. We also analyze model scaling, finding consistent improvements up to 1B parameters, beyond which performance plateaus. Additionally, age-related ASR and speaker verification analysis highlights the limitations of proprietary models like Whisper, emphasizing the need for open-data models for reliable child speech research. All investigations are conducted using ESPnet, and our publicly available benchmark provides insights into training strategies for robust child speech processing.

Descriptors: Child Language, Artificial Intelligence, Natural Language Processing, Automation, Research Problems, Benchmarking, Evaluation Methods, Comparative Analysis, Bias, Models

View on ERIC