Wise Mouth

WiseMouth is a real-time conversation intelligence layer that listens to conversations through AssemblyAI streaming speech-to-text and surfaces observable language signals with supporting evidence. It identifies patterns such as uncertainty, hedged agreement, commitments, changed commitments, repeated concerns, unresolved disagreements, and unanswered questions. WiseMouth does not claim to read emotions, detect lies, profile personalities, or determine what someone “really” thinks. Instead, its interface uses objective language such as “uncertainty language detected” rather than making claims like “Sarah is nervous.” The core principle of WiseMouth is evidence first. Every detected signal includes its type, speaker, verbatim evidence, timestamp, confidence level, topic, and, for temporal signals, the earlier statement it relates to. Temporal intelligence is the key differentiator of the system because WiseMouth can connect statements made at different points in a conversation, such as linking “I’ll handle the deployment” at 00:01 with “I don’t think I’ll have enough time” at 00:28. This allows the system to surface meaningful changes in conversation state rather than treating every statement as an isolated event. WiseMouth uses a hybrid detection engine. Deterministic rules first nominate potential signals cheaply and consistently, after which an LLM can refine or validate those candidates while the conversation state engine connects related statements over time. The system is also designed to degrade gracefully so that the core demonstration remains functional even when some services are unavailable. If microphone access is unavailable, users can type statements or run a prepared demonstration. If an LLM API key is unavailable, WiseMouth can fall back to rules-only detection, and if the database becomes unavailable, the live conversation interface can continue operating without persistence.