The voice becomes a window into cognitive decline before symptoms appear.
En los laboratorios de la Universidad de Alicante y el instituto Isabial, investigadores han enseñado a una máquina a escuchar lo que la voz humana revela antes de que la mente lo sepa: los primeros indicios del Alzheimer. Con una grabación de apenas cuatro minutos y una precisión cercana al 90%, esta plataforma de inteligencia artificial convierte el habla cotidiana en una ventana clínica, desafiando la larga tradición de diagnósticos tardíos que llegan cuando el daño ya es irreversible. En un mundo donde el acceso a la medicina especializada sigue siendo un privilegio, esta herramienta propone que la detección temprana no debería depender de la geografía ni del dinero.
- El Alzheimer avanza en silencio durante años, y cuando la mayoría recibe un diagnóstico, el deterioro cognitivo ya ha dejado huellas profundas e irreversibles.
- Los investigadores identificaron que la voz delata lo que aún no se ha nombrado: el tono que se aplana, las pausas que se alargan, la gramática que se simplifica sin que el hablante lo advierta.
- Con solo 223 grabaciones de voluntarios como base de entrenamiento, el equipo construyó modelos de aprendizaje profundo capaces de distinguir el habla sana del deterioro cognitivo incipiente.
- La herramienta se ejecuta en una aplicación móvil sin equipos costosos ni procedimientos invasivos, pensada expresamente para entornos con recursos limitados.
- La plataforma atraviesa ahora la fase de validación clínica, el umbral que separará el hallazgo de laboratorio de su uso real en consultas y hogares.
Un equipo de investigadores de la Universidad de Alicante y el instituto de investigación sanitaria Isabial ha desarrollado una plataforma de inteligencia artificial capaz de detectar señales tempranas del Alzheimer a partir de una grabación de voz de apenas cuatro minutos, con una precisión cercana al 90%. El sistema fue liderado técnicamente por Miguel Ángel Teruel Martínez junto a Ángel Pérez Sempere, con la colaboración de siete investigadores de ambas instituciones.
La plataforma analiza la voz en dos dimensiones simultáneas. Por un lado, examina sus propiedades acústicas: el tono, el ritmo, la intensidad y, sobre todo, las pausas. Las personas en fases iniciales de Alzheimer tienden a hablar de forma más monótona y a detenerse con mayor frecuencia mientras buscan palabras. Por otro lado, el sistema convierte el habla en texto y estudia los patrones lingüísticos: la complejidad gramatical, la estructura de las frases y la frecuencia de tropiezos verbales. Como explica el propio Teruel Martínez, la herramienta se formula dos preguntas ante cada muestra: qué se dice y cómo se dice.
El modelo fue entrenado con 223 grabaciones de voluntarios, algunos con cognición normal y otros con demencia ya diagnosticada. Sobre esa base, los investigadores aplicaron procesamiento de lenguaje natural y aprendizaje profundo para que el sistema aprendiera a reconocer las huellas acústicas y lingüísticas del deterioro cognitivo.
Lo que distingue a esta herramienta es su vocación de accesibilidad. Funciona como una aplicación móvil que no requiere equipos especializados ni procedimientos invasivos: el usuario simplemente graba su voz leyendo un texto, narrando una historia o respondiendo preguntas estándar. Está diseñada para operar en entornos con recursos limitados, donde las barreras diagnósticas han sido históricamente más altas.
Actualmente en fase de validación clínica, la plataforma deberá confirmar fuera del laboratorio los resultados obtenidos. Si lo logra, podría transformar el calendario del diagnóstico del Alzheimer, permitiendo intervenciones tempranas cuando aún hay más función cerebral que preservar. Además del instrumento diagnóstico, el equipo ha generado un conjunto de datos que sostendrá investigaciones futuras, extendiendo el alcance de su trabajo mucho más allá de esta primera aplicación.
A team of researchers at the University of Alicante and the Isabial health research institute has completed work on a system that listens to the human voice and detects the early signs of Alzheimer's disease. The platform achieves nearly 90 percent accuracy in identifying the condition from a single four-minute recording, making it one of the most efficient screening tools yet developed for a disease that typically goes undiagnosed until significant cognitive damage has already occurred.
The system works by analyzing voice in two distinct ways. It listens to the acoustic properties—the pitch, the loudness, the rhythm of speech, and crucially, the pauses between words. People in the early stages of Alzheimer's tend to speak differently: their tone flattens, their speech becomes more monotonous, and they pause more frequently as they search for words. But the platform also converts speech to text and examines the linguistic patterns themselves. Researchers have found that Alzheimer's patients show recurring structural changes in how they construct sentences, a decline in the complexity of their grammar, and an increase in verbal stumbles and errors. Miguel Ángel Teruel Martínez, a computer science researcher at the University of Alicante who led the technical development alongside Ángel Pérez Sempere from Isabial, explains that the system essentially asks two questions of every voice sample: what is being said, and how is it being said?
The research team built their system on a foundation of 223 voice recordings collected from volunteers—some with normal cognition, others already diagnosed with dementia. This database became the training ground for the artificial intelligence models that now power the platform. The researchers used advanced natural language processing and deep learning techniques to teach the system to recognize the subtle acoustic and linguistic fingerprints of cognitive decline. The work involved contributions from seven researchers across both institutions, each bringing expertise in different aspects of voice analysis and machine learning.
What makes this approach particularly valuable is its accessibility. The platform runs on a simple mobile application that users can operate from home or in a clinical setting. There are no expensive machines, no invasive procedures, no need for specialized equipment. A person simply opens the app and records themselves reading text, telling a story, or answering standard questions. The recordings are then processed to extract the acoustic and linguistic features, which are fed into the trained models for evaluation. The system is designed to work in resource-limited settings, removing barriers that have historically prevented early screening in populations without access to expensive diagnostic infrastructure.
The platform is currently in the clinical validation phase, the stage where researchers test it against real-world cases to confirm that the accuracy rates hold up outside the laboratory. Once it passes this validation, it will move toward actual clinical use. The implications are significant. Alzheimer's disease progresses silently in its early stages, and by the time most people receive a diagnosis, irreversible damage has already accumulated. A tool that can flag cognitive decline from a four-minute voice sample could shift the entire timeline of detection, allowing interventions to begin while there is still more brain function to preserve. The researchers have created not just a diagnostic tool, but a dataset that will support future research into the disease, ensuring that their work extends far beyond this single application.
Notable Quotes
We listen to the voice as pure audio—checking for more pauses, flatter or more monotonous tone—and we also convert speech to text to search for linguistic structures that recur in Alzheimer's patients.— Miguel Ángel Teruel Martínez, University of Alicante researcher
The Hearth Conversation Another angle on the story
Why voice, specifically? There must be dozens of biomarkers for Alzheimer's.
Voice is everywhere. You don't need a blood test or a brain scan. A person can record themselves at home on their phone. That changes who gets screened.
But how reliable is it, really? A person might sound different just because they're tired or stressed.
That's the learning part. The system was trained on 223 people—enough variation to learn what Alzheimer's looks like versus what a bad day sounds like. The 90 percent accuracy suggests it's learned the difference.
What exactly changes in the voice? Is it just slower speech?
It's more subtle. The pitch flattens, pauses get longer, grammar becomes simpler. The system catches all of it at once—the acoustic signature and the linguistic one together. That combination is harder to fake or misinterpret.
So it's in validation now. What happens if it works?
It becomes a screening tool that could reach millions of people who would never get an expensive diagnostic workup. Early detection means early intervention, which might slow the disease before too much is lost.