M1
Talk share
How much of the active discussion time the instructor and the students each spoke, as a percentage.
Classroom observation
Multimodal Utterance Recognition × Reflection
MUR² turns a recorded class discussion into plain measurements of participation: who spoke, for how long, and how long the room waited before someone answered. Then it helps you reflect on them.
MUR² observes. MUR² measures. MUR² helps educators reflect. It does not grade teachers.
Participation, in numbers you can explain. They come from ordinary, unit-tested code. A language model never does the arithmetic; it only helps put the results into words.
M1
How much of the active discussion time the instructor and the students each spoke, as a percentage.
M2
How many different anonymous voices took a real turn: at least 2.0 seconds of speech, and you can change that threshold.
M3
Whether a few voices carry the room: a Gini score, plus the share held by the top four speakers.
M4
Turns that pass from one student straight to another. This is the strongest sign of real discussion.
M5
How long the room waited after a question before someone responded. Groups with fewer than five samples are left out, not guessed.
M6
Brush a stretch of the timeline and every measure is recomputed for just that window.
Every number can be explained
Every number in a report is built to be explainable. When the audio is too unclear to trust a measure, MUR² leaves that measure out instead of guessing.
Follow a recording from classroom audio to a report you can review.
01
Start with a recording of a class discussion.
02
Speech models estimate who spoke when, and what was said.
03
MUR² turns that raw speech into turns, questions and pauses.
04
Deterministic, unit-tested code computes every number.
05
You predict before class, then compare your prediction with what was measured.
06
Take one question into the next session.
When a screen is shared, MUR² reads the slides on your own device with Apple’s Vision framework, then deletes the screenshots. Only the slide titles, their text and the times are kept.
Those slide words go to the speech recognizer as a glossary, so subject terms come out spelled the way your slides spell them.
The report then shows the slides beside the Murmur Timeline: one lane per speaker, layers you can switch on and off, and a brush to zoom into any stretch. Low-confidence stretches are hatched rather than tinted, and colour is never the only cue.
Slide 1
Introduction
Slide 2
Heteroscedasticity
Slide 3
Worked example
Glossary from slide text
…the residuals show heteroscedasticity, so the usual assumption fails…
MUR² observes. MUR² measures. MUR² helps educators reflect. It does not grade teachers.
There is always a report: with no API keys at all, a built-in fallback writes the plain-language summary. When a language model is used, it only narrates numbers that code has already computed.
01
Before class, commit to a prediction of how the discussion will go.
02
Before the results appear, estimate what the numbers will show.
03
See your prediction, your estimate and the measured result side by side.
04
Leave with a single question to carry into the next session.
A classroom is not a data source to be mined. These limits are part of the design, not settings to hunt for.
What it does
What it will never do
Version 1.4.0. Built and tested, and honest about what is still unproven.
In place and tested
Not yet validated
What MUR² reports has not yet been checked against a hand-labelled class. Treat every number as a measurement with unknown error until it has.
The checks for speakers, talk time, questions and wait time are built and ready. Telling voices apart from across a real classroom is hard, and no public benchmark exists for it.
Inspired by Loquat (Lachko Hristov, RTI International, CIES 2024). MUR² is an independent implementation and is not affiliated with RTI.