Pedro Benevenuto Valadares
Research Data Scientist · Speech & Voice AI · Campinas, Brazil
I build real-time voice agents, fine-tune speech models for noisy and low-bandwidth audio, and optimize inference on CPU and GPU. Currently at Itaú's applied AI research institute (ICTi) and pursuing an M.S. in Electrical Engineering at UNICAMP.
Recent repos
A voice agent that runs (almost) entirely in the browser: speech recognition, LLM and speech synthesis on a laptop GPU through WebGPU. 1.4–2.1 s voice-to-voice on a GTX 1050 notebook GPU.
HTML labdemowikikernelsPublications
-
Neural Speech Enhancement for Conversational AI over Narrowband Telephony: A Benchmark in Brazilian Portuguese
BRACIS 2026, Springer LNAI -
Synthetic Speech Detection in Brazilian Portuguese through Accent-Related Features
IEEE SLT 2026 (accepted) -
Extracting Accent Features in Spoken Brazilian Portuguese without Sociolinguistic Labels
SBrT 2026 (accepted) -
PCA-based Channel Estimation for MIMO Communications
SBrT 2022 -
Song Emotion Recognition: A Performance Comparison Between Audio Features and Artificial Neural Networks
REIC, vol. 20, no. 4, 2022
Research
Generates room impulse responses from text prompts in real time, combining signal processing with generative models. The goal is to give AI agents a sense of the acoustic space around them.
Individualized headphone equalization
Personalized headphone equalization for binaural (3D) audio, using regularized inverse filters with automatic parameter selection to avoid audible peaks. Advised by Prof. Bruno Sanches Masiero at UNICAMP. Grant 20/10195-7.
Experience
Itaú ICTi (applied AI research)
Data Scientist II, Conversational AI & Voice
Research on real-time conversational voice AI: low-latency audio streaming, parameter-efficient fine-tuning of speech recognition models for noisy, narrowband audio, and inference optimization on GPU and CPU for many concurrent audio streams.
Nokia
NPO AI/Automation Engineer (Machine Learning & AI team)
Computer vision for edge and telecom use cases: segmentation, classification and object detection.
Ericsson
Data Analyst
Time series forecasting for anomaly detection and root cause analysis across global products.
Education
M.S. Electrical Engineering, UNICAMP
Advisor: Prof. Bruno Sanches Masiero