[All work](/work?type=work)

# Conversational Timing Lab

I wanted to know whether a pause means somebody has finished or is just thinking, so I built a lab around that difference.

[Source code](https://github.com/Slaymish/conversational-timing-lab)

Independent research. Dataset adapters, causal labels, baseline evaluation and audio-model pipeline.

![Two E0 plots: turn-end probability falling from 0.633 to 0.435 as the silence lengthens, and false interruption rates of 0.53 to 1.00 against a 5% budget.](https://cdn.sanity.io/images/qnuj1c4o/production/c2cb4e43560ca2f157eac3affdfe652a7e8174b3-1600x900.png?w=1280\&q=80\&auto=format)

Turn completion and response onset are separate targets here, even though both look like silence in the audio. The labels come from who spoke next, and observations are limited to what a listener would actually have at that moment, with splits by conversation so the same meeting can't leak across training and validation.

The baseline run covered 100,004 boundaries from 171 AMI meetings. The audio-model run checks the pipeline end to end, and it isn't a trained predictor. Pause length on its own separated the two cases badly.

## State

A reproducible baseline and a training pipeline. Pause length turned out to carry less than it looks like it should.

[All work](/work?type=work)