Lab notes
From the lab
A running record of what we are testing, publishing, and learning as projects move from the bench into the world.
Changing phones should not change the answer
Two phones can listen to the same machine and produce noticeably different spectra. Some of what the model hears comes from the microphone, not the machine.
Tashfain Ahmed, Mohammadali Saffary, and I have been working on cross-device spectrum correction so a model trained with one device can travel more reliably to another. Tashfain is presenting the work at ACM GoodIT 2026.
A score is not the whole answer
Machine-opinion scores usually return a single number, even when the model is unsure. ConformalMOS adds an interval around that estimate, making the uncertainty visible instead of burying it.
Kehinde Elelu leads this work with Mohammadali Saffary, Tashfain Ahmed, Simeon Babatunde, Ebuka Okpala, and me. We are presenting it at Interspeech 2026. The goal is straightforward: a decision-support system should know when to ask for a closer look.
We made the test harder on purpose
A model can look impressive while quietly learning the room, the microphone, or the recording day. Our new dataset makes those shortcuts harder. It includes shop vacuums, a household vacuum, and an orbital sander recorded across sessions, locations, microphone positions, phones, and external microphones.
That variation gives us a better test of whether a diagnostic model has learned something about the machine itself.
The bench is only the beginning
Kyle Foster and Tejas Agrawal joined Paremeswar Nair in undergraduate research on acoustic diagnostics. They are taking the same basic question—what can we learn by listening to a machine?—to portable generators, air compressors, ordinary phones, and less-than-ordinary recording conditions.
Every new machine brings different operating states, background noise, and sensor-placement problems. That is where the useful research begins.
Read the work
The publications page has the papers, datasets, and links behind these notes.