Projects, research, and results
A look at recent research and the kind of work I've led. Client details are kept confidential.
Perceptual Metric Benchmark
The question: can objective and AI-based audio quality metrics predict what trained listeners actually hear?
What I did: compared 34 metrics, 17 of them AI-based, with expert MUSHRA listener ratings in Python. Then I trained a machine-learning model that combines several metrics into one score.
The result: the combined model reached r = 0.89 on trials it had never seen, versus 0.74 for the best single metric. Where the metrics fall short is documented, too.
Why it matters: listening tests are slow and expensive. Knowing when a metric is good enough, and when it isn't, saves time without risking quality.
Correlation with expert listener ratings (unseen trials)
34 metrics tested · 17 AI-based · Pearson r
Selected work at Xperi/DTS
Sound Profile framework
Directed research on a framework of 10+ perceptual attributes that translated target listener experiences into measurable success criteria tied to objective metrics.
Virtual:X perceptual targets
Defined pass thresholds for virtual width and height that products had to meet, and advised field engineering on customer tuning.
Certification standards
Turned perceptual findings into standards, acceptance criteria, and regression tests, cutting certification development time 30%. In use for 5+ years.
BS.1116 listening rooms
Renovated listening rooms to full BS.1116 compliance for reliable, repeatable test results.
Products shaped through listening tests
- DTS:X
- Virtual:X
- Headphone:X
- Clear Dialogue
- Automotive audio proof of concept
Have a product that needs to sound right?
Or a team that needs a perceptual audio lead? Either way, let's talk.