NeuroVoice

NeuroVoice is a multimodal machine learning platform for neurological screening from speech and language, built during my time as founding AI developer at Rice University. It’s since moved from research prototype to a production system piloted by state health agencies.

the problem

Conditions like Parkinson’s, ALS, and cognitive decline show up in speech - articulation, rhythm, prosody, word-finding - often before they’re caught by standard clinical screening. But cognitive screening tools that rely on in-person administration don’t scale to a state health agency trying to screen tens of thousands of people a year.

what I built

As the founding AI developer, I built the ML core of the platform: models that take raw speech and language input and output neurological risk signals, plus the pipelines to train and validate them at scale.

class NeuroVoiceModel(nn.Module):
    def __init__(self):
        # Pretrained speech encoder for rich acoustic representations
        self.encoder = WhisperEncoder.from_pretrained("base")

        # Multimodal head combining acoustic + linguistic features
        self.classifier = nn.Sequential(
            nn.Linear(512, 256),
            nn.ReLU(),
            nn.Dropout(0.3),
            nn.Linear(256, num_conditions)
        )

training data and validation

The models were trained on 3,000+ hours (12TB+) of clinical and public-health audio - a scale that matters, because neurological speech biomarkers are subtle and demographically variable. Against baseline cognitive screening methods, the platform achieved a 25–30% accuracy improvement.

deployment

The platform is now piloted with state health agencies across 8 US states, supporting roughly 50,000+ screenings per year. That production deployment, plus the traction from the pilots, helped the effort raise $150K+ in seed funding.

what I learned

Building for a health-agency pilot is a different discipline than building for a benchmark. Every model decision had to be explainable to a non-ML stakeholder, and the data pipeline had to handle real-world audio - background noise, accents, recording quality - not curated clinical recordings.