Hi, I’m Hugo.

I help teams bring speech recognition into their products and work through the audio problems that show up in practice.

My work spans both integrating transcription into products and measuring how well it performs.

I built and open-sourced gladia-normalization, a multilingual text-normalization library that Gladia uses internally for speech-to-text benchmarking and competitive evaluations.

I worked on a conversational engine with dynamic agent orchestration, real-time code execution through MCP, and interchangeable speech-recognition (ASR) and text-to-speech (TTS) components.

I'm interested in the tooling around LLMs: the integrations and decisions that turn a model into a useful system.

Speech recognition has to hold up outside clean recordings. I worked on noise augmentation for a multilingual speech-to-text training pipeline and measured its effect on word error rate (WER) in difficult audio.

That work keeps me interested in audio pipelines, training, evaluation, and real-world robustness.

At Gladia, I built a cross-platform dictation application with Tauri. It brings real-time transcription into a desktop workflow and inserts the resulting text across the system.

The code is open if you want to see how it works.

Production audio is often unlabeled, while labeling customer calls is costly, slow, and constrained by privacy. Clean public datasets rarely reflect noisy telephony or far-field recordings.

NoiseKit adds codecs, reverb, noise, and clipping to labeled audio. This makes it possible to compare models by WER under realistic deployment conditions.