A study investigates conversational timing as a controllable variable for training synthetic multi-speaker conversations in automatic speech recognition (ASR). Researchers parameterized pause and overlap distributions using an exponential-tilting family and explored the resulting four-dimensional space via Latin hypercube sampling and multi-objective Bayesian optimization.
- Simulated timing configurations were used to train ASR systems and evaluate performance on a Hungarian dialogue corpus using cpWER and cpCER metrics.
- Higher exposure to speech overlap correlates with lower word error rates, while longer and more variable gaps are associated with higher error rates.
- Bayesian optimization provided modest aggregate improvements but primarily served an analytical role by revealing the specific trade-off between overlap and gap timing.
The findings suggest that realistic simulation should be complemented by task-relevant diagnostics of overlap, gap, and timing-variability profiles to improve training data utility.