Teleprompter

Captures audio from active meetings and converts speech to text in real-time using on-device speech recognition (likely leveraging Web Audio API or native OS audio capture). The system maintains a rolling context window of recent transcribed speech to understand conversation flow and speaker intent, enabling contextually-aware suggestion generation without sending raw audio to external servers.

Uses a lightweight language model (likely a quantized or distilled model for on-device execution) to analyze the current meeting context and generate charismatic, relevant quote suggestions in real-time. The system takes the recent transcription history and speaker intent as input, then produces suggestions ranked by relevance and rhetorical impact, enabling speakers to inject compelling language without interrupting their flow.

Implements a multi-factor ranking system that scores generated suggestions based on relevance to current conversation topic, alignment with speaker intent, rhetorical appropriateness, and estimated charisma impact. Uses heuristics or learned scoring functions to filter low-quality suggestions and surface the most contextually-appropriate options, preventing overwhelming the user with irrelevant recommendations.

Provides a unified interface to capture audio from multiple meeting platforms (Zoom, Google Meet, Microsoft Teams, etc.) by abstracting platform-specific audio APIs and system-level audio routing. Handles permission negotiation, audio format normalization, and fallback mechanisms to ensure consistent transcription input regardless of the underlying meeting application.

Displays generated suggestions in a non-intrusive UI overlay (likely a floating panel or sidebar) that appears in real-time without blocking the meeting view. Implements fast dismissal and acceptance mechanisms (keyboard shortcuts, click-to-insert) to minimize disruption to the speaker's flow, with latency-optimized rendering to ensure suggestions appear within 1-2 seconds of generation.

Ensures all processing (speech recognition, LLM inference, suggestion ranking) occurs entirely on the user's device without transmitting audio, transcripts, or suggestions to external servers. Implements local model loading, in-memory processing, and optional encrypted local storage for conversation history, providing end-to-end privacy guarantees without requiring trust in third-party services.

Maintains a bounded buffer of recent conversation history (likely 5-15 minutes of transcribed speech) that serves as context for suggestion generation and relevance scoring. Implements efficient memory management to keep only recent utterances in active memory while optionally archiving older history to disk, enabling the system to understand conversation flow without unbounded memory growth.

Analyzes recent conversation context to classify the current speaker's intent (e.g., persuading, explaining, asking for feedback) and detect the primary topic being discussed. Uses lightweight classification models or heuristic rules to tag utterances with intent and topic labels, enabling suggestion generation to be tailored to the specific communicative goal rather than generating generic suggestions.

Provides mechanisms to accept a suggestion and automatically insert it into the meeting context (e.g., typing into chat, inserting into speaker notes, or copying to clipboard). Handles platform-specific insertion methods (Zoom chat, Teams message, Google Meet chat) and implements undo/edit capabilities to allow users to modify suggestions before committing them.