EchoForge

EchoForge is a Python CLI that turns audio into structured local knowledge artifacts.

The pipeline covers three steps:

1. Fetch or stage audio from Feishu Minutes or a local file
2. Submit the audio to a supported ASR provider for transcription and meeting understanding
3. Render the results into Obsidian Markdown notes

Status

The Python implementation under src/echoforge/ is the current pipeline. It has been verified end-to-end with Tingwu + R2 transit, and it also supports Doubao ASR.

Recent additions:

• Dual ASR provider support: Switch between Tingwu and Doubao via ECHOFORGE_UNDERSTANDING_PROVIDER
• Gemini post-processing: Optional summarization, chapter extraction, Q&A, and action items generated from the transcript markdown
• Long-audio segmentation: Files longer than 119 minutes are automatically split at silence points and processed in parallel
• State persistence: All runs are tracked in outputs/runs.json with list-runs and inspect-run commands

EchoForge also works with Feishu Minutes exports that already contain transcript and summary data. When feishu_minutes_sync exports a minute from the web page, EchoForge can render the standardized artifacts directly without sending the audio to a third-party ASR provider again.

One important integration constraint comes from the current provider APIs:

• Offline transcription tasks require a public FileUrl
• Local file paths are staged into outputs/, and the provider still needs an externally reachable HTTP or HTTPS URL

That means process-file is implemented, but you must provide --media-url unless your source already exposes a downloadable URL.

Obsidian Output Structure

EchoForge uses the dual-file structure settled on 2026-04-16:

• Summary note: meetings/{date}-{title}.md
• Transcript note: meetings/Transcripts/{date}-{title}-transcript.md
• Vault index: EchoForge Index.md

The summary note links into transcript block anchors, so different upstream providers can share one rendering format.

Standard Artifacts

EchoForge renders from a normalized artifact layer. Current standard files are:

• transcription.json
• chapters.json
• summarization.json
• meeting_assistance.json

Different upstream sources can map into the same shape:

• Tingwu returns all four categories directly
• Doubao returns transcription, chapters, summarization, and information extraction URLs
• Feishu Minutes web export now produces:
  • transcript.vtt
  • transcription.json
  • summarization.json
  • chapters.json

This keeps the final Obsidian output stable across providers.

Installation

python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"

Configuration

Copy .env.example to .env and fill in at least:

# ASR provider selection: tingwu | doubao
ECHOFORGE_UNDERSTANDING_PROVIDER=tingwu

# Tingwu
TINGWU_ACCESS_KEY_ID=...
TINGWU_ACCESS_KEY_SECRET=***
TINGWU_APP_KEY=...

# Doubao (when provider = doubao)
DOUBAO_APP_KEY=...
DOUBAO_ACCESS_KEY=...

# Obsidian
OBSIDIAN_VAULT_PATH=/path/to/vault

# Optional: Gemini post-processing
GEMINI_API_KEY=...
GEMINI_BASE_URL=https://generativelanguage.googleapis.com
GEMINI_MODEL=gemini-2.0-flash
GEMINI_ENABLE_SUMMARY=true

Optional Feishu settings:

FEISHU_MINUTES_SYNC_BIN=feishu-minutes-sync
FEISHU_MINUTES_SYNC_EXPORTS_DIR=./exports

Feishu Minutes web export workflow:

1. Use feishu_minutes_sync export-minute --token <minute_token> --fetch-mode web
2. Exported files land under exports/<minute_token>/
3. Render the transcript directly with render-transcript, or feed the normalized JSON files into the full EchoForge renderer

Commands

# Process audio
python -m echoforge process-feishu <minute_token>
python -m echoforge process-file ./recording.ogg --media-url https://example.com/recording.ogg

# Re-render an existing run
python -m echoforge render <run_id>

# Render a standalone transcript JSON
python -m echoforge render-transcript ./transcription.json --title "导入转写" --output-vault ~/Obsidian/vault

# State management
python -m echoforge list-runs
python -m echoforge list-runs --status failed
python -m echoforge inspect-run <run_id>

Render a transcript from Feishu Minutes standardized export:

python -m echoforge render-transcript \
  ../feishu_minutes_sync/exports/<minute_token>/transcription.json \
  --title "会议标题" \
  --output-vault ~/Obsidian/vault \
  --note-name imported-transcript \
  --source-label "Feishu Minutes WEBVTT"

Use python -m echoforge --help for the full CLI.

Layout

EchoForge/
├── config/
├── outputs/
│   └── runs.json
├── src/echoforge/
└── tests/

Typical run artifacts:

outputs/runs/run_<timestamp>_<hash>/
├── run.json
├── media.ogg
└── results/
    ├── transcription.json
    ├── chapters.json
    ├── summarization.json
    └── meeting_assistance.json

Testing

pytest