Zero-Shot Parkinson's Disease Detection from Speech: Comparing Large Audio and Language Models

A new study examines whether handcrafted acoustic features or raw audio waveforms are more effective for zero-shot Parkinson’s disease detection across languages, finding that input modality significantly shapes performance [1]. Large audio and language models have recently shown zero-shot reasoning capabilities across multiple domains, but their application to Parkinson’s disease detection from speech has not been systematically compared by input type [1]. The study, posted to arXiv on 24 May 2026, evaluates two approaches: handcrafted acoustic features extracted from speech recordings and analyzed by a general-purpose large language model, and direct waveform input processed by audio-capable models [1]. Experiments were conducted on Parkinson’s disease speech datasets spanning four languages [1]. Results indicate that performance varies not only by input modality but also by the specific speech task and language [1]. Handcrafted acoustic features delivered more stable results in a low-resource language, Bengali, while raw audio input produced gains that depended on the particular dataset [1]. The authors note that these findings underscore the importance of input modality when designing zero-shot detection systems for neurological conditions [1]. Parkinson’s disease is a progressive neurodegenerative disorder that affects motor control and often impairs speech production. Researchers have increasingly explored vocal biomarkers as a non-invasive screening tool, leveraging advances in machine learning and signal processing. The current study extends this line of inquiry by testing whether large pre-trained models can detect the disease without task-specific fine-tuning, a capability known as zero-shot learning [1]. The work appears amid broader scientific efforts to apply artificial intelligence to health diagnostics. While the research bundle does not contain additional studies on Parkinson’s detection, the arXiv preprint itself provides a self-contained comparison of modality effects [2]. The authors do not report specific accuracy figures in the abstract, and no external validation data are cited in the available materials [1]. Future research will likely need to examine whether these modality-dependent patterns hold across more languages and clinical settings. The preprint’s findings suggest that developers of speech-based screening tools should consider both the type of audio input and the target language when selecting model architectures [1].