i Imran E Kibria | Portfolio
Profile Picture

Imran E Kibria

I am a fourth-year CSE PhD student in the ASPIRE Lab at the Ohio State University. I specialize in speech quality assessment, speech enhancement, and building neural networks that generalize across diverse acoustic conditions. My research spans deep learning methods including architecture design, domain generalization, and small-scale modeling for edge applications.

My research has been featured at ICASSP, Interspeech and EMBC conferences. I enjoy connecting with researchers and collaborating to advance speech and audio products through novel ideas and impactful innovations.

News & Updates

  • Awarded summer grant by the Center of Cognitive and Brain Sciences to research the cognitive perception of speech quality.
  • Our work on improving domain generalization of speech quality networks has been accepted at the IEEE ICASSP'26 Conference.
  • Our work on small scale modeling of speech quality for edge applications has been published at the ISCA Interspeech'25 Conference.
  • Awarded merit-based fellowship "CSE Scarlet & Gray Award" by the Ohio State University for 5 years.
  • Our work on automation of color assesssment for urine test analysis has been published at the IEEE EMBC'22 Conference.

Research Projects

Loss Landscape

One Model for All Domains: Zero-Shot Speech Quality Assessment

  • Analyzed factors influencing domain shift in MOS-label datasets.
  • Improved generalization by optimizing model weights to a flat minima.
  • Investigated loss landscape and dataset distributions in latent space.
  • Reported improvement over AdamW optimizer on 8 of 12 evaluation sets.
Paper Paper Icon | Code GitHub Icon | Poster Abstract Icon
AttentiveMOS

Lightweight AI for Perceptual Speech Quality on the Edge

  • Designed MOS prediction network for edge applications using 86K params.
  • Analyzed transformer variants & positional encodings as design choices.
  • Countered subjective noise in MOS ratings using a distillation strategy.
  • Wav2vec-level performance with 100× fewer parameters on SOMOS dataset.
Paper Paper Icon | Code GitHub Icon | Presentation Presentation Icon
Urinox 10

Smartphone-Based Urine Test Strip Analysis for At-Home Diagnostics

  • Algorithm for automated urine test strip analysis using smartphone camera.
  • Lighting & chromatic calibration using multivariate regression scheme.
  • Color matching with reference chart in different colors spaces.
  • 92% in-house accuracy; low-cost alternative to laboratory analyzers.
Paper Paper Icon | Code GitHub Icon | Presentation Presentation Icon

Publications