---
title: Podify - AI Podcast Generator
emoji: 🎙️
colorFrom: indigo
colorTo: purple
sdk: gradio
sdk_version: 5.27.0
app_file: app.py
python_version: "3.10"
hardware: zero-a10g
suggested_hardware: zero-a10g
pinned: false
short_description: Research a topic and turn it into a voiced podcast
tags:
  - track:backyard
  - achievement:offgrid
  - achievement:offbrand
  - achievement:fieldnotes
---

![Podify banner](assets/readme-banner.png)

# 🎙️ Podify — AI Podcast Generator

Turn any topic into a finished, voiced podcast in two phases:

1. **Content** — research agents (LangGraph) use a HuggingFace-hosted LLM plus live
   DuckDuckGo web search to research the topic and write a speaker-tagged script.
2. **Audio** — the self-hosted **Fish Audio / OpenAudio S1-mini** model speaks the
   script, with selectable preset voices and zero-shot **voice cloning** from an
   uploaded clip or a live mic recording.

Everything runs inside this single Gradio Space; the TTS model runs on **ZeroGPU**.

## Architecture

```
Topic ─▶ LangGraph: plan ─▶ DDG search ─▶ outline ─▶ write ─▶ Script
Script ─▶ Fish Audio S1-mini (@spaces.GPU): per-line synth ─▶ stitched podcast WAV
```

- `app.py` — Gradio Blocks UI (two tabs) wiring both phases.
- `research/` — `llm.py` (HF Inference client), `search.py` (DuckDuckGo), `graph.py`
  (LangGraph research graph).
- `tts/` — `engine.py` (model load + GPU synthesis + multi-speaker stitching),
  `voices.py` (preset voice registry).

## Configuration

Set these as **Space secrets / variables** (Settings → Variables and secrets):

| Name        | Required | Purpose                                                        |
|-------------|----------|----------------------------------------------------------------|
| `HF_TOKEN`  | ✅       | LLM inference (Inference Providers) + model download.          |
| `LLM_MODEL` | optional | Override the content LLM (default `Qwen/Qwen2.5-14B-Instruct`, <32B). |
| `TTS_MODEL_REPO` | optional | Override the TTS model repo (default `fishaudio/openaudio-s1-mini`). |

**ZeroGPU** requires a HuggingFace **PRO** account on the Space owner.

## Run locally

```bash
pip install -r requirements.txt
export HF_TOKEN=hf_xxx          # PowerShell: $env:HF_TOKEN="hf_xxx"
python app.py
```

Phase 1 (research + script) runs on CPU. Phase 2 (TTS) needs a GPU and the
`fish-speech` package; on CPU-only machines the UI loads but synthesis is disabled.

## Models Used
- **Qwen/Qwen2.5-7B-Instruct** For Research and Script Generation
- **fishaudio/openaudio-s1-mini 0.5b** For Audio Generation

## Deploy to a Space

```bash
huggingface-cli login
huggingface-cli upload <user>/podify . --repo-type=space
# or: git push to the Space remote (preset .wav files tracked via Git LFS)
```

## Credits / assets

- **Voice samples** (`tts/voices/`): derived from [CMU ARCTIC](http://festvox.org/cmu_arctic/)
  (free for research and commercial use). Rebuild with `scripts/build_voice_samples.py`.
- **Background-music loops** (`tts/music_loops/`): [FreePD](https://freepd.com/) by Kevin
  MacLeod — 100% public domain (CC0). Rebuild with `scripts/build_music_loops.py`.
  A procedural numpy fallback in `tts/music.py` is used if the loops are absent.

## Contributots
- **nvipin63**
- **jayaspjacob**


#backyard-ai
- Blog: [Article](https://huggingface.co/blog/build-small-hackathon/podify)
- Social Media Post: [Post](https://substack.com/@nvipin63/note/c-276881572?r=637t58&utm_source=notes-share-action&utm_medium=web)
- Demo: [Video](https://youtu.be/DRVf_Q8IoOI)