Meeting Transcript Ingest

Meeting Transcript Ingest imports Google Meet transcripts, Lark/Feishu Minutes transcripts, and local audio transcriptions into an Obsidian material library. It is designed for crypto security teams that need meeting transcripts to become durable source notes, not forgotten chat artifacts.

The default output target is a ZeroDrift-style Obsidian vault:

~/Documents/Obsidian Vault/ZeroDrift Material Library/sources/meetings

What It Does

• Imports Google Meet transcript entries through the Google Meet REST API.
• Imports Lark/Feishu Minutes transcripts through the Minutes OpenAPI.
• Searches Lark/Feishu Minutes when a user token with search permission is available.
• Transcribes local audio with mlx_whisper on Apple Silicon.
• Falls back to OpenAI gpt-4o-transcribe through a bundled transcribe CLI.
• Optionally generates a Chinese-first ZeroDrift meeting brief with OpenAI.
• Preserves crypto/security vocabulary such as EVM, Solana, CLMM, oracle, MEV, governance, multisig, TVL, and proof-of-exploit.

Install

cd ~/meeting-transcript-ingest
python3 -m venv .venv
. .venv/bin/activate
pip install -r requirements.txt

Install the local Apple Silicon transcription fallback:

pip install mlx-whisper

ffmpeg must also be installed for most audio workflows:

brew install ffmpeg

Configure

Copy the example env file and fill only the credentials you need:

cp .env.example .env

Do not commit .env or files under state/*.json; they are ignored by git.

Google Meet

Create a Google OAuth desktop client, save it as:

state/google-oauth-client.json

The importer requests:

https://www.googleapis.com/auth/meetings.space.readonly

Import by conference record:

python scripts/meeting_ingest.py google-import \
  --conference-record "conferenceRecords/CONFERENCE_RECORD_ID" \
  --title "Customer security review sync"

Or import a known transcript resource:

python scripts/meeting_ingest.py google-import \
  --transcript-name "conferenceRecords/CONFERENCE_RECORD_ID/transcripts/TRANSCRIPT_ID" \
  --title "Customer security review sync"

Lark / Feishu

By default, the importer prefers an installed lark-cli and falls back to direct OpenAPI environment tokens. This lets the automatic service reuse your existing Lark CLI profile without storing a Lark token in .env.

For a single Minutes transcript:

python scripts/meeting_ingest.py lark-import \
  "https://example.feishu.cn/minutes/obcnq3b9jl72l83w4f149w9c" \
  --region feishu \
  --title "Weekly security sync"

For global Lark:

python scripts/meeting_ingest.py lark-import \
  "https://example.larksuite.com/minutes/obcnq3b9jl72l83w4f149w9c" \
  --region lark

Search Minutes:

python scripts/meeting_ingest.py lark-search \
  --region feishu \
  --query "周会" \
  --start-time "2026-05-27T00:00:00+08:00" \
  --end-time "2026-05-27T23:59:59+08:00"

Search requires a Lark/Feishu user access token with the Minutes search scope. Export can use a user or tenant token if the app has access to the Minutes file.

Use LARK_BACKEND=cli to require lark-cli, or LARK_BACKEND=api to force direct OpenAPI tokens.

Local Audio Fallback

Use MLX Whisper locally:

python scripts/meeting_ingest.py transcribe-audio meeting.m4a \
  --engine mlx \
  --title "Crypto customer call"

Use OpenAI for higher accuracy with a crypto vocabulary prompt:

python scripts/meeting_ingest.py transcribe-audio meeting.m4a \
  --engine openai \
  --openai-transcribe-model gpt-4o-transcribe \
  --title "Crypto customer call"

Add --summarize when OPENAI_API_KEY is available:

python scripts/meeting_ingest.py lark-import "MINUTES_URL_OR_TOKEN" \
  --region feishu \
  --summarize \
  --write-internal

--write-internal also writes the same note into:

~/Documents/Obsidian Vault/ZeroDrift Wiki/raw/internal

Automatic Execution

The tool can run automatically on macOS through launchd. The installed job runs one import cycle every hour.

Install the automatic job:

cd ~/meeting-transcript-ingest
./scripts/install_launchd.sh

Check status:

./scripts/launchd_status.sh

Uninstall:

./scripts/uninstall_launchd.sh

Logs are written to:

logs/daemon.out.log
logs/daemon.err.log

The automatic runner executes:

python scripts/meeting_ingest.py --env-file .env daemon-run --once

Each hourly cycle does three things:

1. Searches Lark/Feishu Minutes for recent transcript records and imports unseen transcripts.
2. Best-effort checks Google Meet transcripts when an existing Google OAuth token is present.
3. Processes files under inbox/.

The daemon deduplicates imported items through:

state/daemon-state.json

Inbox Automation

Drop transcript or audio files into:

inbox/

Supported transcript files:

.txt .md .vtt .srt .tsv

Supported audio files:

.mp3 .mp4 .mpeg .mpga .m4a .wav .webm

Audio files are transcribed with mlx_whisper by default. Set AUDIO_ENGINE=openai in .env to use gpt-4o-transcribe.

One-Shot Daemon Test

Run one cycle without writing anything:

python scripts/meeting_ingest.py daemon-run --once --dry-run

Run one real cycle:

python scripts/meeting_ingest.py daemon-run --once

Useful .env settings:

LARK_BACKEND=auto
LARK_CLI=lark-cli
LARK_CLI_PROFILE=
LARK_CLI_AS=user
LARK_REGION=feishu
LARK_QUERY=
LOOKBACK_HOURS=72
LARK_PAGE_SIZE=20
GOOGLE_AUTO=1
GOOGLE_PAGE_SIZE=20
MEETING_INBOX_DIR="./inbox"
AUDIO_ENGINE=mlx
MLX_WHISPER_MODEL="mlx-community/whisper-large-v3-turbo"
OPENAI_TRANSCRIBE_MODEL=gpt-4o-transcribe

Automation Roadmap

The current automatic layer is a local poller/scheduler. The next layer is event-driven automation:

1. Subscribe to Google Meet transcript fileGenerated events through Google Workspace Events API and Pub/Sub.
2. Replace Lark/Feishu polling with webhooks when the deployed app has the required permissions.
3. Trigger google-import or lark-import directly from each platform event.
4. Keep the local inbox fallback for manually exported transcripts and audio recordings.

Security Notes

• This repo should never contain OAuth tokens, .env, transcript exports, or Obsidian vault content.
• OAuth tokens are stored under state/ and ignored by git.
• Public examples use fake IDs and tokens.
• For sensitive meetings, prefer official platform transcripts or local MLX transcription before uploading audio to a cloud API.