Preserve the previous agent turn for voice routing
For reply-conditioned voice routing, preserve the immediately preceding agent turn
Short caller replies such as “yes,” a bare number, or a city are not self-describing. A training exporter must preserve the exact preceding agent utterance and bind it causally to the caller audio; treating that agent turn as leakage changes an easy conversational classification task into an ambiguous audio-only task. First prove a call-disjoint transcript-conditioned ceiling, then replace caller text with audio under the identical contract. This does not remove label noise or rare-class support requirements.
Applies when: a voice agent selects tools from terse replies whose meaning depends on its last question. Keywords: voice routing previous agent turn causal context audio classification LiveKit