Hedy

Captures live audio streams from video conference platforms (Zoom, Teams, Google Meet) and converts speech to text in real-time using cloud-based ASR (automatic speech recognition) with speaker identification. The system maintains a rolling buffer of audio chunks, processes them through a speech recognition API, and tags utterances with speaker identities by analyzing audio characteristics and meeting participant metadata. Transcription is streamed to the UI as it completes, enabling live note-taking without post-call processing delays.

Processes completed transcripts through a multi-stage NLP pipeline: first, a summarization model (likely fine-tuned T5 or BART) condenses the full transcript into 2-3 paragraph executive summary; second, a named entity recognition (NER) + dependency parsing layer identifies action items, decisions, and owners by detecting imperative verb phrases and linking them to speaker identities; third, a topic segmentation model breaks the meeting into logical sections (agenda items, discussions, decisions). The system uses extractive + abstractive hybrid summarization to preserve exact quotes while generating coherent prose.

Indexes all meeting transcripts using full-text search (likely Elasticsearch or similar) combined with semantic search via embedding vectors (sentence transformers or OpenAI embeddings). When a user searches, the system performs hybrid retrieval: keyword matching for exact phrase queries (e.g., 'budget approved $50k') and semantic similarity for conceptual queries (e.g., 'what did we decide about pricing?'). Results are ranked by relevance and returned with context snippets showing the speaker, timestamp, and surrounding dialogue. Supports filtering by date range, attendees, and meeting type.

Stores meeting recordings (audio or video) in cloud object storage (likely AWS S3 or similar) with automatic transcoding to multiple bitrates for adaptive streaming. The playback interface synchronizes the transcript timeline with video/audio playback: clicking a transcript line seeks the recording to that timestamp, and the current playback position highlights the corresponding transcript line in real-time. Supports variable playback speed (0.5x to 2x) and speaker filtering (hide/show specific speakers' audio). Recordings are encrypted at rest and access-controlled via user permissions.

Integrates with calendar systems (Google Calendar, Outlook, Zoom, Teams) via OAuth 2.0 to detect scheduled meetings and automatically join video calls. When a meeting starts, Hedy's bot joins the call (as a participant or via platform API), captures audio, and begins transcription without requiring manual user action. The system extracts meeting metadata (title, attendees, duration) from calendar events and associates it with the transcript. Supports recurring meetings and handles timezone conversions for global teams.

Aggregates data across all meetings to generate analytics: meeting frequency trends, average meeting duration, attendee participation rates, decision velocity (time from discussion to decision), and topic frequency analysis. The dashboard uses time-series visualization (line charts for trends), heatmaps for attendee participation patterns, and word clouds for common topics. Data is computed via batch jobs (daily or weekly aggregation) rather than real-time, and results are cached for fast dashboard load times. Supports filtering by date range, attendee, and meeting type.

Provides a web-based editor for users to manually correct transcription errors (typos, misheard words, speaker labels) after the meeting. Changes are tracked with version history: each edit creates a new version with timestamp and user attribution, allowing rollback to previous versions. The editor uses a diff-based approach to highlight changes between versions. Corrections can be applied to individual words, phrases, or entire speaker turns. The system supports bulk find-and-replace for common errors (e.g., correcting a company name misspelled throughout the transcript).

Exports transcripts in multiple formats: plain text (.txt), Microsoft Word (.docx), PDF, JSON (structured with speaker labels and timestamps), SRT (subtitle format for video sync), and CSV (for spreadsheet analysis). The export pipeline handles format-specific requirements: PDF includes formatting and page breaks, Word documents preserve speaker labels and timestamps in a table, JSON maintains full metadata, and SRT generates subtitle timing for video players. Users can customize export options (include/exclude timestamps, speaker labels, summary, action items) before generation.

Implements role-based access control (RBAC) for meetings and transcripts: Owner (full access), Editor (can edit transcript and share), Viewer (read-only access), and Commenter (can view and add comments). Sharing is done via shareable links with optional password protection and expiration dates (e.g., link expires after 7 days). Access logs track who viewed the transcript and when. Supports team-level permissions (all team members can access team meetings) and individual-level permissions (specific users granted access). Integrates with SSO (single sign-on) for enterprise authentication.