IISURI LAB

RESEARCH · INAOE · 2004

Infant-cry recognition with intelligent hardware.

Electronic implementation of a hybrid neural network for classifying infant acoustic signals through a parallel FPGA architecture.

Israel Alejandro Suaste RivasMaster of Science, Electronics · National Institute of Astrophysics, Optics and Electronics
ProblemClassify acoustic cry patterns and possible pathological conditions.
MethodLPC + PCA + fuzzy relational hybrid neural network.
ImplementationCeloxica RC100 FPGA with parallel Handel-C design.

Research summary

Infant cry is one of the primary channels a baby uses to communicate. The thesis studies how acoustic characteristics can distinguish three classes: normal crying, crying from infants with deafness, and crying associated with asphyxia.

The proposal combines linear prediction coefficients, principal component analysis for dimensionality reduction, and a hybrid neural network. The classifier is first evaluated through simulation and then implemented in reconfigurable hardware, using parallel execution to reduce processing time.

2,268one-second audio segments analyzed in the initial dataset.
94.6078%correct classification reported for Group A with triangular membership functions.
3.2 msestimated processing time per vector in the FPGA implementation.
English translation of the master’s thesis cover
Original document. Thesis submitted at INAOE, Tonantzintla, Puebla, December 2004.

System architecture

The flow takes an audio sample, digitizes it, extracts features, reduces the input vector, and sends a decision through a trained network. The architecture is divided into stages that can execute in parallel inside the FPGA.

InputMono audio digitization at 16 kHz, with frequencies below 8 kHz retained through filtering.
FeaturesLinear prediction coefficients representing the acoustic behavior of the vocal tract and cry signal.
ReductionPrincipal component analysis reduces the coefficient volume while preserving useful classifier information.
ClassificationFuzzy relational neural network with membership functions and three outputs: normal, deafness, and asphyxia.
HardwareParallel implementation on the RC100 FPGA, with memory, keyboard, and monitor result output.
Infant-cry recognition flow diagram
Recognition flow. Acoustic processing, feature reduction, and classification.
Schematic of the neural network implemented on the FPGA
Electronic implementation. Schematic of the fuzzy relational neural network on the FPGA.

Results and conclusions

The best Group A result used seven triangular membership functions and achieved 94.6078% correct classification. In Group B, the same configuration reached 92.292%. The hardware implementation reproduced the simulation behavior and used FPGA parallelism to shorten processing time.

Classification results graphs from the thesis
Evaluation. Accuracy comparison across sample groups and membership-function configurations.
Why it still matters

The thesis anticipates engineering patterns that remain relevant in AI systems: feature extraction, dimensionality reduction, hybrid classification, parallel acceleration, reproducible evaluation, and reusable design.

Complete document

This page is an editorial summary for easier reading. The complete document preserves the original text, tables, figures, and appendices.

Open complete thesis ↗Back to credentials