The current arXiv version was revised on October 20, 2026, and the paper is listed as accepted to IEEE SLT 2026.
The authors study prior-turn context in multi-turn speech recognition. They report that, after supervised multi-turn training, conversational context mainly improves recognition of contextual entities. Raw prior-turn audio embeddings grow with conversation length, so the method replaces each prior audio sequence with a fixed number of learned latent tokens while keeping the corresponding transcripts explicitly available.
On in-domain and out-of-domain test sets, the authors say the compression-enabled model recovers part of the gain from raw-context conditioning with a smaller prior-turn audio footprint. The abstract does not provide absolute metrics, implementation cost, code or data availability, or evidence that the method is ready for production use.