Reproduce “Speaker-Aware Simulation Improves Conversational Speech Recognition” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Máté Gedeon, Péter Mihajlik
Researchers adapted and extended the Speaker-Aware Simulated Conversations (SASC) framework for Hungarian, demonstrating its utility in improving conversational Automatic Speech Recognition (ASR) performance on a real dataset. Their duration-conditioned SASC (C-SASC) variant further refined temporal modeling, yielding modest but systematic gains, particularly for character-level metrics, when statistics matched the simulated material.
Key insights
- Augmenting ASR training data with speaker-aware simulated conversations consistently improves recognition performance on real conversational benchmarks.
- Incorporating utterance-duration-conditioned pause modeling (C-SASC) refines temporal realism, leading to modest but systematic improvements, especially in character-level ASR metrics, with effectiveness dependent on the match between statistical source and simulated data.
- The utility of Room Impulse Response (RIR) augmentation in simulated conversational data depends on the evaluation context, with observed degradation when applied to ASR models evaluated on high-quality studio recordings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.