Automatic dysarthria detection holds promise for clinical screening in settings with limited specialist access, yet two systemic problems hinder progress: speaker-leaking evaluation protocols may inflate results, and a systematic cross-lingual transfer analysis within a unified framework is absent. We present a leave-two-speakers-out (LTSO) protocol that eliminates speaker leakage. Additionally, we present a 4×4 transfer matrix for the TORGO (English, henceforth EN), AISHELL-Dys (Chinese Mandarin, henceforth ZH), HDS (Russian, henceforth RU), and EasyCall (Italian, henceforth IT) datasets. Finally, we present an analysis of factors that may drive transfer effectiveness. Key findings are as follows: (1) a random split inflates accuracy on EN from 65 F_1=0.850 ; (2) ZH appears to be the best transfer source, seemingly due to speaker diversity rather than dysarthria severity, with macro F_1=0.955 when tested on RU; it is also the optimal single source for both target languages, and combining EN and ZH does not improve transfer over ZH alone; (3) speech material type appears to matter more than linguistic proximity – an F_1=0.570 when tested on IT may be caused by the voice-command format of EasyCall; (4) simple target-domain feature normalisation outperforms DANN ( F_1=0.791 vs F_1=0.661 when trained on EN and tested on RU).
Modern expressive Text-to-Speech (TTS) systems for adult speech can generate highly expressive speech when provided with a sufficient amount of training data. However, expressive TTS systems for children are often restricted and have been under-explored to date. Child TTS systems’ challenges arise from the scarcity of training datasets, difficulties in accessing them publicly, and various issues associated with these datasets, including noisy data and indiscernible speech. This work introduces an expressive Multi-Style Child TTS in low-resource settings. Our model stands as the first expressive TTS model designed specifically for children. It is a style-based generative model based on the StyleTTS architecture. Our model is trained on an expressive ChildTinyTalks (CTT) dataset, comprising 2 h of speech, covering four speaking styles. We verify the framework through subjective and objective evaluations and demonstrate that Multi-Style Child TTS outperforms other baseline models.
Dysarthria, a motor speech disorder caused by neurological impairments, substantially reduces speech intelligibility and poses a major challenge for conventional Automatic Speech Recognition (ASR) systems. Variability in articulation, timing, and prosody often produces ambiguous acoustic patterns, making direct transcription unreliable. To address this problem, this paper presents VocalPlan-AV, a two-stage audio-visual framework for dysarthric ASR based on a reverse engineering perspective. Instead of directly mapping distorted speech to text, the proposed method first reconstructs an intermediate structured representation, referred to as a vocal plan, which captures the intended phoneme sequence, temporal structure, and selected prosodic and articulatory cues from synchronised audio and visual inputs. By jointly modelling respective acoustic and visible articulatory information, such as lip movements, the framework leverages complementary cross-modal cues when the speech signal is unclear. The reconstructed vocal plan is then encoded and used to guide a task-specific language model trained from scratch for transcript generation, rather than relying on any existing pre-trained large language model. This two-stage design aims to separate linguistic intent from motor execution distortion and thereby improve robustness to slurred, slow, and imprecise speech. Experiments on the UA-Speech corpus show that the proposed framework can recover meaningful linguistic content from dysarthric speech while providing an interpretable intermediate representation for assistive speech technologies. Compared with the baseline systems on UA-Speech, the proposed VocalPlan-guided audio-visual model achieves a WER of 16.88
The paper presents a study on the use of large language models (LLMs) in safety-critical medical helplines for supporting parents of young children. We propose an LLM-driven conversational system designed as a pre-triage assistant for medical staff rather than an autonomous advisor. The system conducts an initial dialogue with callers, collects structured symptom and context information, and produces a summary for the medical expert. It integrates automatic speech recognition, speech synthesis, dialogue management, and summarization. We evaluate multiple LLMs (Gemma 3, EuroLLM, Salamandra), fine-tuned on a newly created Serbian-English dataset of 150 simulated pediatric helpline dialogues, on the task of generating the next response in such a dialogue. LLMs are evaluated under different prompting strategies (zero-shot, few-shot, prompt variants) and languages (English, Serbian). Results show that improved prompting and increased model capacity generally lead to better performance, with few-shot prompting yielding the best overall results and increasing BERTScore from 0.701 to 0.739 in the strongest configuration. We further observe consistent cross-lingual generalization across English and Serbian, with varying sensitivity to prompting style across models. However, we also find that standard metrics correlate poorly with clinical correctness, which motivates domain-specific evaluation for more reliable assessment of system performance.
