Project · Product Hub

Aurora — Spec: Audio-to-Events Pipeline

High confidenceconceptedited by Cairni · 방금 · AIv4

Overview

The Audio-to-Events Pipeline is the foundational workstream of the Aurora project. It is responsible for ingesting meeting audio recordings and automatically extracting two structured outputs: calendar events and action items. This pipeline underpins all downstream features, including the Shared Team Calendar and Slack Notifications. Aurora — Kickoff Brief.md

Requirements

#	Requirement	Notes
R1	Accept meeting recordings in audio format as input	MVP scope Aurora — Kickoff Brief.md
R2	Auto-extract calendar events from audio	Core extraction capability; uses ElevenLabs for transcription with diarization Feature Spec — Audio to Events.md
R3	Auto-extract action items from audio	Core extraction capability; extracts task, owner, and due date Feature Spec — Audio to Events.md
R4	Output feeds the shared team calendar (week view)	Consumed by Calendar spec Aurora — Kickoff Brief.md
R5	Output feeds Slack notifications for assigned action items	Consumed by Slack Notifications spec Aurora — Kickoff Brief.md
R6	Resolve relative dates with confidence flags	Ambiguous dates (e.g., "next Tuesday") resolved relative to meeting date; low-confidence dates flagged for user confirmation Feature Spec — Audio to Events.md
R7	User review & confirmation step	Extracted items are reviewed by the user before being committed to the calendar and Slack Feature Spec — Audio to Events.md

Behavior

1.Ingestion — A user or system submits an audio recording of a meeting.
2.Transcription — Audio is processed using ElevenLabs with diarization (speaker identification) for high-accuracy transcription, particularly for overlapping speech. Feature Spec — Audio to Events.md
3.Extraction — An LLM processes the transcription and identifies candidate:

Events (title, datetime, attendees) to populate the team calendar.
Action items (task, owner, due date) to be tracked and notified.

4.Date Resolution — Relative date references are resolved against the meeting date; low-confidence resolutions are flagged for user confirmation. Feature Spec — Audio to Events.md
5.User Review — Extracted events and action items are presented to the user for confirmation before commit.
6.Distribution — Confirmed events are pushed to the Shared Team Calendar; confirmed action items trigger Slack notifications for the assigned team member. Feature Spec — Audio to Events.md

Data Model

Event

Field	Source	Notes
Event title	transcript	required Feature Spec — Audio to Events.md
Datetime	transcript + resolution	low-confidence dates flagged Feature Spec — Audio to Events.md
Attendees	transcript	inferred from diarization

Action Item

Field	Source	Notes
Task description	transcript	required
Owner	transcript	defaults to speaker if not explicit Feature Spec — Audio to Events.md
Due date	transcript + resolution	low-confidence dates flagged

Process Flow

Edge Cases

Feature Spec — Audio to Events.md

Overlapping speech — When multiple speakers talk simultaneously, the LLM relies on ElevenLabs' diarization output to separate contributions by speaker. Low-confidence extractions are flagged for review.
No explicit date — When a task or event is mentioned without a specific date, the item is left unscheduled and the user is prompted to provide a due date during the review step.

Ownership

Role	Person
Backend Owner	Dev Patel
PM	Mina Park
Engineering Lead	Jun Lee

Aurora — Kickoff Brief.md