ssdavidai/ambient-alfred/omi-transcripts

Omi Ambient Transcripts

You have access to ambient audio transcripts captured by an Omi wearable device. These are continuously recorded, transcribed, and organized into conversations.

Transcript Location

Transcripts are saved as JSON files in a date-organized directory:

$ALFRED_TRANSCRIPTS_DIR/
  2026-02-27/
    14-30-00_device1.json
    14-35-22_device1.json
    ...

Each JSON file contains:

• uid: device identifier
• text: transcribed text
• language: detected language code
• timestamp: ISO 8601 timestamp
• audio_duration_seconds: total audio length
• speech_duration_seconds: actual speech length

Conversations

Segments are grouped into conversations by the pipeline:

• 10-minute silence gap = new conversation
• Conversations with < 30 words are filtered as noise
• Each conversation is written to the vault inbox as markdown with YAML frontmatter

Vault Inbox Files

Processed conversations appear in $ALFRED_VAULT_INBOX_DIR/ as:

omi-2026-02-27-1430-1445.md

With YAML frontmatter containing: type, date, time, duration, segments, languages, words.

Command Detection

If the agent's name (default: "Alfred") is mentioned in a conversation, the pipeline uses an LLM classifier to determine if it's a direct command. If so, a subagent is spawned to execute it.

Working with Transcripts

When asked about conversations or what was discussed:

1. Check the vault inbox directory for recent markdown files
2. Read the transcript content to understand what was discussed
3. Transcripts may be in multiple languages — provide translations if needed
4. Pay attention to timestamps and duration for context