I am a speech-language pathologist, acoustic phonetician, and neural network engineer developing next-generation clinical speech technologies. I have expertise across the full lifecycle of clinical AI development: speech representations, neural network training, clinical trialing, and scalable deployment.

My work bridges clinical practice, speech science, and AI to develop theoretically motivated and clinically validated tools that help clinicians extend the reach of best-practice speech therapy.

I am currently a post-doctoral fellow in the Speech Communication Lab at the University of Maryland, mentored by Dr. Carol Espy-Wilson, where I am training in acoustic-to-articulatory speech inversion and speech modeling. My training and research have been supported by an NIH T32 award.

Research Themes

Clinically Tested AI Speech Therapy

My research develops clinical AI systems grounded in speech-language pathology best practice. I am the lead inventor on the patent-pending PERCEPT AI system, a neural network that predicts clinician judgment of speech correctness during therapy. My work with Jonathan Preston demonstrated that speech production improves with ChainingAI, a therapy platform delivering motor-based intervention with automated feedback that allows learners to practice evidence-based Speech Motor Chaining at home. As a clinician-engineer, I collaborate on multiple clinical trials for AI speech therapies and on establishing responsible standards for how clinical AI should be developed and evaluated.

Automated Kinematic Speech Analyses

My postdoctoral research investigates acoustic-to-articulatory speech inversion, a technique that infers vocal tract articulatory kinematics from audio alone. My research has shown that inferred tract variables differentiate subtypes of speech sound errors in children with single-sound speech disorders and quantify coordination differences in childhood apraxia of speech. My recent work has further demonstrated that gradient perceptual ratings of speech correctness predict articulatory proximity to the target sound, providing evidence that clinician ratings capture underlying articulatory distance and not only intermediate acoustic cues. This work directly motivates my ongoing research toward a personalized AI biofeedback system that delivers vocal tract gesture feedback inferred from audio alone.

Ultrasound Biofeedback and Acoustic Biofeedback for Speech Therapy

My colleagues and I have shown that motor learning is enhanced during speech therapy that includes biofeedback, such as ultrasound biofeedback and acoustic biofeedback. We have shown that biofeedback is effective, particularly when delivered through intensive schedules, even for children who have not responded to traditional speech therapy. A recent large-scale randomized controlled trial with Tara McAllister, Jonathan Preston, and Elaine Hitchcock found that biofeedback produces significantly faster speech improvement than motor-based treatment during the acquisition phase.

Speech Profiles of Speech Sound Disorders in Children

I have investigated how different types of speech sound disorders manifest across articulatory, perceptual, and acoustic dimensions, as well as how to automate the measurement of these dimensions in collaboration with Katie Hustad and Visar Berisha.

Child Speech Data and Research

I have curated the PERCEPT Corpora, which are currently the largest open-access speech corpora focused on speech sound disorders in American English. I am also the sole US-based co-founder of the Interspeech Child Speech Science and Speech Technology special sessions, and played a role in introducing child speech topic areas to the main program.

Automated Language Sample Analysis

My ongoing postdoctoral research examines the use of automatic speech recognition to automate clinical language sample analysis. My work has shown that automated language sample metrics outperform manual metrics in predicting language disorder status in children, reducing the burden of manual transcription and scoring while improving diagnostic sensitivity.

The tools I use to address these research themes include: Azure DevOps, Amazon AWS, Git, Google Speech API, HTCondor/Cluster Computing, Linux, MATLAB, Montreal Forced Aligner, NLTK, Phon, Praat, PyTorch/CUDA, R, REDCap API, Regex, RESTful APIs, SAS, Stanza/CoreNLP, and several custom tools I’ve developed in Python