Data

Overview

Dataset for the EverMemOS competition — 832 real conversation events captured by the Spiro wearable band. All transcripts have been speaker-normalized and converted to structured formats for downstream analysis.

Files

File	Description	Size
basic_events_79ef7f17.json	Raw dataset with embedded speaker mappings	13 MB, 832 events
gcf_all.json	Merged GroupChatFormat output (72 conversations, 15K messages)	5.2 MB
last100_events.json	Subset of recent events for quick testing	1.4 MB
demo_audio.mp3	Audio sample from the Spiro wearable	6.5 MB
demo_output.json	Sample analysis output	1.7 KB

Data Format

Event Structure

Each event in basic_events_79ef7f17.json is a JSON object with two top-level keys:

{
  "meta": {
    "user_id":          "User ID (UUID)",
    "basic_event_id":   "Event ID (UUID)",
    "basic_start_time": "Start time (Unix epoch seconds)",
    "basic_end_time":   "End time (Unix epoch seconds)"
  },
  "object": {
    "basic_transcript": "Normalized transcript text"
  }
}

meta Fields

Field	Type	Description
user_id	string	User unique identifier (UUID)
basic_event_id	string	Event unique identifier (UUID)
basic_start_time	number	Event start time, Unix epoch seconds
basic_end_time	number	Event end time, Unix epoch seconds

object Fields

Field	Type	Description
basic_transcript	string	Speaker-normalized transcript (see format below)

Fragment Structure

Transcripts are organized into fragments, each representing a contiguous segment of conversation:

[Fragment N: YYYY-MM-DD HH:MM - YYYY-MM-DD HH:MM]
Title: Conversation title
Type: career, social, home, ...

[Speaker1]: Utterance text
[User]: Utterance text
[Colleague/Friend]: Utterance text

Speaker Label Normalization

Speaker labels are normalized by scripts/normalize_speakers.py into the following categories:

Label Type	Format	Description
Primary user	[User]	The person wearing the recording device
Generic speaker	[Speaker1], [Speaker2], ...	Numbered by order of appearance; numbering resets per event
Generic + gender	[Speaker1/M], [Speaker2/F]	With gender annotation
Named role	[Colleague/Friend], [Partner], [Interviewer], etc.	Original role description preserved

Dialogue Line Format

All dialogue lines follow a consistent format (no timestamps):

[Speaker Label]: Utterance content

Non-Dialogue Lines

Transcripts may contain the following non-dialogue lines, which are automatically skipped during pipeline parsing:

• Fragment header: [Fragment N: ...]
• Title / Type: Title: ... / Type: ...
• Metadata: [Full Transcript and Summary], etc.
• Environment notes: [Quiet environment] [Clear audio], etc.
• Passive media: Passive media, transcript content omitted

Pipeline Flow

Spiro Wearable Band
    │
    ▼
basic_events_79ef7f17.json   ← 832 raw events with speaker mappings
    │
    ├─► normalize_speakers.py ← Standardize speaker labels
    │
    ▼
gcf_all.json                 ← Merged GroupChatFormat for EverMemOS
    │
    ▼
EverMemOS API (localhost:1995)

1. Capture: The Spiro band records conversations and produces raw event transcripts.
2. Normalize: scripts/normalize_speakers.py standardizes all speaker labels into a consistent format.
3. Convert: The pipeline converts normalized events into GroupChatFormat (gcf_all.json), which merges related fragments into conversations.
4. Ingest: GCF data is posted to the EverMemOS API for memory storage and retrieval.