PodSight：用 Whisper 与 Gemini 自动转录并发布台湾财经播客

Turn podcast audio into structured, searchable intelligence — automatically.

PodSight is a fully automated pipeline that transcribes, summarizes, and publishes Taiwan finance podcasts. New episodes are detected via RSS, transcribed with Whisper, summarized by Gemini AI, published to a beautiful static site, and pushed to Telegram — all running on GitHub Actions, twice daily, zero human intervention.

Live site: podsight.tw | Telegram: @podsight

How It Works

RSS Feed --> Download MP3 --> Whisper Transcription --> Gemini Summary --> Static Site --> Telegram

Every day at 10 AM and 7 PM (Taiwan time), GitHub Actions:

1. Fetches RSS feeds for all configured podcasts
2. Detects new episodes by comparing RSS against existing summaries
3. Downloads audio and transcribes via Groq Whisper API
4. Generates AI summaries with Gemini — structured with key topics, stock mentions, market insights
5. Builds a static site with individual episode pages (deployed to Vercel)
6. Pushes to Telegram after verifying the Vercel deployment is live

Public Site

The generated site features a dark-themed, mobile-friendly design with episode cards and detailed summary pages.

Each episode page includes a one-line summary, structured topic breakdowns, stock mentions, and actionable insights — all extracted by AI from the full transcript.

Telegram Channel

Summaries are automatically pushed to the @podsight Telegram channel with direct links to the full summary page.

Web Dashboard

A local admin dashboard for managing the pipeline, browsing episodes, editing summaries, and searching across all transcripts.

Podcasts Currently Tracked

Adding a new podcast is a single entry in podcasts.yaml — no code changes needed.

Quick Start

Prerequisites

• Python 3.10+
• FFmpeg (brew install ffmpeg on macOS)

Setup

git clone https://github.com/jazzpujols34/podsight.git
cd gooaye_pipeline

python -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Create a .env file:

GROQ_API_KEY=your_groq_api_key        # Whisper transcription
GEMINI_API_KEY=your_gemini_api_key     # AI summarization
TELEGRAM_BOT_TOKEN=your_bot_token      # Optional: Telegram push
TELEGRAM_CHAT_ID=your_chat_id          # Optional: Telegram push

Run the Full Pipeline

# Auto-detect and process all new episodes across all podcasts
python src/pipeline/auto_pipeline.py

Run Individual Steps

# Target a specific podcast
PODCAST=gooaye python src/pipeline/01_parse_rss.py      # Fetch RSS
PODCAST=gooaye python src/pipeline/02_download_audio.py  # Download MP3s
PODCAST=gooaye python src/pipeline/03_transcribe.py      # Transcribe
PODCAST=gooaye python src/pipeline/04_summarize.py       # Summarize
PODCAST=gooaye python src/pipeline/05_generate_social.py # Generate social drafts

Web Dashboard

python src/server.py --port 3501
# Open http://localhost:3501

The dashboard lets you browse episodes, view transcripts/summaries, edit drafts, search across all content, and trigger pipeline runs manually.

Adding a New Podcast

Edit podcasts.yaml:

podcasts:
  my_podcast:
    name: "My Podcast"
    slug: my_podcast
    host: "Host Name"
    rss_url: "https://example.com/feed.xml"
    language: zh
    episode_pattern: 'EP(\d+)'  # Regex to extract episode number (null for date-based)
    max_episodes: 30             # Limit for daily podcasts

That's it. The pipeline will pick it up on the next run.

Project Structure

gooaye_pipeline/
├── src/
│   ├── config.py                  # Podcast configs, shared utilities
│   ├── server.py                  # FastAPI web dashboard (port 3501)
│   ├── pipeline/
│   │   ├── 01_parse_rss.py        # RSS feed -> episodes.json
│   │   ├── 02_download_audio.py   # Download MP3 files
│   │   ├── 03_transcribe.py       # Audio -> text (Groq Whisper)
│   │   ├── 04_summarize.py        # Transcript -> summary (Gemini)
│   │   ├── 05_generate_social.py  # Summary -> social media drafts
│   │   ├── auto_pipeline.py       # Orchestrator: runs all steps
│   │   ├── generate_public_site.py # Static HTML generator
│   │   └── push_telegram_batch.py # Telegram channel publisher
│   └── social/
│       ├── formatters/            # Platform-specific formatters (Telegram, Twitter, etc.)
│       └── publishers/            # Platform publish adapters
├── ui/                            # Web dashboard frontend (SPA)
├── public-site/                   # Generated static site (deployed to Vercel)
├── data/{podcast}/                # Per-podcast data
│   ├── episodes.json              # RSS metadata
│   ├── transcripts/*.txt          # Transcription output
│   ├── summaries/*_summary.txt    # AI-generated summaries
│   └── social_drafts/             # Telegram/social media drafts
├── podcasts.yaml                  # Podcast configuration
├── requirements.txt               # Python dependencies
└── .github/workflows/
    └── auto-pipeline.yml          # Scheduled CI/CD (twice daily)

Deployment

The production setup uses three services:

GitHub Actions secrets needed: GROQ_API_KEY, GEMINI_API_KEY, TELEGRAM_BOT_TOKEN, TELEGRAM_CHAT_ID

Tech Stack

• Transcription: Groq Whisper API (whisper-large-v3)
• Summarization: Google Gemini with custom prompts per podcast
• Backend: Python, FastAPI
• Frontend: Vanilla HTML/CSS/JS SPA (no framework)
• Static Site: Custom HTML generator with Lucide icons, dark theme
• CI/CD: GitHub Actions
• Hosting: Vercel (static site)
• Distribution: Telegram Bot API

License

MIT — Pipeline code is open source. Podcast content belongs to respective creators.