Dr. Hugo Arnoldo
Mitre Hernández

A hybrid MFCC pipeline on ESP32-S3 for joint anger detection and speaker identification.

Voice-based emotion recognition has attracted attention for early-warning monitoring in vulnerable households, yet most systems run on phones or PCs — raising cost and stripping speaker identity. We propose a compact hybrid framework that extracts affective and identity-related cues from a single MFCC front-end, deployed on an ESP32-S3 microcontroller for domestic-violence early-warning scenarios.

Two complementary INT8 branches share the same 20×63 MFCC matrix: MicroLightCNN, a LightCNN-derived network for binary anger detection, and MLP-XiEmbedding, an MLP over a deterministic four-statistic Xi-Vector pooling (μ, σ, max, min over 20 Mel bands). Aggregating both cues in one capture window widens perceptual capacity without breaking the memory budget.

On the Emotional Speech Dataset (ESD) and an in-house Mexican-Spanish household corpus, the pair consistently outperforms MCU-deployable variants of DS-CNN, MatchboxNet, ECAPA-TDNN and CAM++ on the joint accuracy–memory–latency frontier. A sliding-window vote of three frames suppresses spurious activations under realistic acoustic variability; the device emits contextual alerts of the form "Dad angry — 63 %" in under 500 ms per 2-second window.

Head-motion recognition for intelligent wheelchair control.

An inclusive artificial-intelligence framework for electric wheelchair control that enables users with severe motor impairments to navigate through eight head-motion commands, including cardinal and diagonal directions. Unlike conventional systems that rely on fixed-threshold detectors and limited command sets, the proposed MSM–LSTM framework models head movements as non-stationary multivariate time series, improving robustness to temporal variability across users.

The system integrates a multi-criteria evaluation protocol for elastic similarity measures, an optimised Proximity Forest 2.0 classifier, and a hybrid MSM–LSTM architecture with confidence-based softmax fusion. Validation was conducted in a Unity-based virtual-reality supermarket using an HTC Vive Pro Eye headset with synchronised data acquisition at 5 Hz.

Results identified MSM, SSDTW, and ERP as the most effective similarity measures, while the proposed classifier achieved perfect F1 scores for cardinal commands and a macro-averaged AUC of 0.91, outperforming a two-layer LSTM baseline. The framework provides a computationally efficient and reproducible solution for accessible head-controlled assistive technologies.

Anxiety and stress recognition with physiological sensors.

We investigate physiological signals and body movements for real-time detection of anxiety and stress, developing a novel signal-data-selection method tailored to microcontrollers that handle small physiological inputs.

In parallel we design deep-learning models optimised to use computational resources efficiently on the same constrained hardware — keeping inference small, fast, and local.

Eye-tracking data collection in VR driving behaviour.

We study visual distraction during driving inside a virtual-reality environment. The project's centrepiece is a novel model that gauges mental effort by analysing real-time changes in pupil size, enabling us to quantify the challenge posed by distracting driving tasks as they unfold.

Webcam-based eye tracker.

Bringing eye-tracking to commodity hardware. We combine computer vision, deep learning, and curated face-eye data collection to develop accurate, low-cost webcam-based eye trackers that bring gaze analytics out of the lab and onto everyday devices.