All projects

Punctuation restoration research

Improving STT output by restoring punctuation from multimodal signals

TebahSoftSep 2022 – Dec 20224 mosUniversity–industry collaboration

Context

A university–industry project on STT output quality, the front end of the counselling analysis pipeline.

Problem

ASR output arrives without punctuation, which hurts readability and drags down every downstream step — summarisation, emotion analysis, and the rest.

How I solved it

  1. Surveyed prior work and confirmed where text-only restoration breaks down
  2. Designed and built a multimodal model that infers from text and audio features together
  3. Validated the PoC through a Streamlit demo

Result

  • F1 score improved 15% over text-only punctuation restoration

Impact

Everything downstream of STT — summarisation, emotion analysis — now receives cleaner input.

Tech

PythonPyTorchStreamlit
NextPervisFirst contract, 7 months in