Buzz-Bot

Podcast player for the AI epoch.

Listen to any podcast, then tap one button and hear it in your language — same voices, different words.

Buzz-Bot runs inside Telegram as a Mini App and uses a cloud GPU pipeline (RunPod Serverless) to transcribe, translate, and re-synthesize every speaker's voice.

To view AI pipeline for podcast dubbing workflow check this repository

---

Features

Feature	Description
RSS subscriptions	Add any podcast by RSS URL; bulk-import via OPML
Podcast search	Search Apple Podcasts directory and subscribe in one tap
Episode inbox	Unified feed of unheard episodes across all subscriptions, with "hide listened" and compact grouping filters
Episode player	Native audio playback inside Telegram with resume-from-position, variable speed (1x/1.5x/2x), skip, and a persistent mini-player
Autoplay	Automatically advance to the next episode when one finishes
Progress tracking	Listening position saved every 5 seconds; offline saves queued and replayed on reconnect
Offline caching	Episode audio downloaded in background; seamless switch to local copy on network loss
Bookmarks	Bookmark episodes with a single tap; search saved episodes
Semantic search	Search your inbox by meaning — episodes re-ranked by vector similarity to your query — details
Topic cloud	Browse episodes by auto-extracted topics, clustered nightly across languages (EN/RU/NL) so synonyms collapse into one tag; tap a tag to filter your library — details
Smart recommendations	Hybrid semantic + collaborative filtering with explainable topic matching — details
Share & send	Share any episode via Telegram's share sheet, or send the audio file directly to your own chat
AI dubbing	Hear any podcast in your language with the original speaker's voice cloned — details
Karaoke subtitles	Live subtitle panel with karaoke-style highlighting; fullscreen transcript; tap any line to seek

---

Tech Stack

Layer	Technology
Language	Crystal >= 1.9
Web server	Kemal
Telegram bot	Tourmaline
Database	PostgreSQL (Neon) via crystal-pg, pgvector for embeddings
Frontend	ClojureScript, re-frame, Reagent
Frontend build	shadow-cljs
Service Worker	Offline audio cache (Range-aware) + offline write queue
Job dispatch	RunPod Serverless API v2
AI pipeline	Demucs, WhisperX, pyannote, Gemini Flash, VoxCPM2 — details
Embeddings	BGE-M3 (1024-dim, multilingual) + KeyBERT topic extraction — details
Audio storage	Cloudflare R2
Deployment	Docker, k3s on Hetzner — details
Ingress / TLS	Traefik v3, cert-manager + Let's Encrypt

---

Quick Start

Prerequisites

• Crystal >= 1.9 + shards
• Node.js >= 18 + npm
• Docker
• PostgreSQL (Neon free tier works)
• Telegram bot token from @BotFather
• Public HTTPS URL for webhooks

1. Clone and install dependencies

git clone https://github.com/yourname/buzz-bot.git
cd buzz-bot
shards install
npm install

2. Configure environment

cp k8s/secret.example.yaml k8s/secret.yaml
# Fill in all values — see docs/DEPLOYMENT.md for variable reference

3. Run migrations

for f in migrations/*.sql; do psql "$DATABASE_URL" -f "$f"; done

4. Build frontend

# Development (live reload)
npx shadow-cljs watch app

# Production
npx shadow-cljs release app

5. Run locally

crystal run src/buzz_bot.cr

Use Cloudflare Tunnel to expose localhost over HTTPS for Telegram webhooks:

# Terminal 1 — ClojureScript watch build
npx shadow-cljs watch app

# Terminal 2 — tunnel + Crystal server
./devrun.sh           # auto-detects named tunnel or quick tunnel
./devrun.sh --quick   # temporary URL, no account needed

---

Documentation

• AI Dubbing Pipeline — how dubbing works, supported languages, progress tracking
• Recommendations — hybrid vector + collaborative filtering, topic extraction & clustering, explainability
• Deployment Guide — k3s setup, environment variables, monitoring
• API Reference — routes, database schema, feature flags

---

License

MIT