Audio–video quality assessment with IISc Bangalore
Predict how people would rate a TV's audio and video quality, so testing doesn't depend on large human studies.
- collaboration
- 18 mo
- Indian Institute of Science, Bangalore
- IISc
- degradation datasets
- Real + synthetic
Context
TV testing needs to know how good the picture and sound look and sound to people, but large human rating studies are slow and expensive.
This collaboration with the Indian Institute of Science (IISc), Bangalore explored AI models that predict perceived audio–video quality for consumer-TV testing pipelines.
What I built
- Capture setups that route a TV's audio and video output into PC-based analysis pipelines through capture cards, for controlled benchmarking and dataset creation.
- Multi-stage dataset pipelines that combine real TV output with synthetically generated audio and video degradations.
- Evaluations of the audio and video quality models proposed by the academic team.
Outcome
- Suggested architectural and training refinements to the academic models, including embedding-space analysis and mono/stereo handling.
- The datasets were used for systematic evaluation, fine-tuning and refinement of the models.