We also tried a semantic layer: transcribing each call with ElevenLabs Scribe and extracting text-based cues (filler words, lexical diversity, speaking-rate variance). Measured head-to-head against our validated model, it didn't add meaningful lift on its own — but rather than drop it, we exposed it as an optional mode, togglable per request from the frontend, so it's available live for cases where the extra semantic signal (or an ElevenLabs-powered attack) is worth the added latency and cost.
Log in or sign up for Devpost to join the conversation.