Cesar Petrescu

Lab · AI

  • In development

SparkTranscript

Romanian speech research and a medical-dictation website

SparkTranscript combines work on Romanian speech recognition with a website for medical dictation using a foot pedal. I evaluate and fine-tune speech models, test them with real microphone input, and build the web experience through which clinics can request a demo.

  • NVIDIA NeMo
  • FastConformer
  • Qwen3-ASR
  • Next.js
  • TypeScript
  • PostgreSQL

Components

Speech research

  • Evaluate and fine-tune NeMo FastConformer and Qwen3-ASR on H100 GPUs.
  • Use real-microphone testing to assess whether the Romanian speech stack is ready for practical deployment.

Website

  • Romanian and English pages explain medical dictation with a foot pedal.
  • Demo requests are stored through a server API backed by PostgreSQL.

Production notes

Live website; speech validation ongoing
The website is live. The speech work is still being validated for production readiness; no accuracy percentage is claimed here.
Product scope
The current offer is transcription. Form autofill, templates and EMR integration remain outside the shipped scope.