AI-Podcast is an intelligent audio content generation platform powered by Multi-Agent collaboration. It transforms topics, documents, or URLs into immersive, podcast-style audio conversations. By leveraging state-of-the-art LLMs and TTS models, it orchestrates a team of AI agents—from planners to directors and voice actors—to produce high-quality, structured, and engaging audio content in real-time.

• 👥 Flexible Cast Formats: Supports Solo (Monologue), Duo (Dialogue), and Multi-person (Roundtable) modes to suit different content styles.
• 🧠 Adaptive Depth Modes:
  • Lite Mode: Quick, concise summaries for rapid consumption.
  • Deep Exploration Mode: In-depth analysis where agents perform parallel web searches and structured outlining for comprehensive coverage.
• 🎭 Custom Personas & Voices: Users can fully customize character personalities (system prompts) and timbre (voice cloning/selection).
• 🗣️ Real-time Interaction: Supports user intervention, allowing you to join the discussion and steer the conversation in real-time.
• 📚 Diverse Inputs: Generate podcasts from a simple Topic, uploaded Documents (PDF/Txt), or URLs.

This project is built upon a robust stack of cutting-edge open-source tools:

• Multi-Agent Framework: AgentScope - Orchestrates the complex interaction between agents.
• Large Language Models (LLM):
  • Qwen/Qwen3-8B
  • tencent/Hunyuan-7B-Instruct
• Text-to-Speech (TTS):
  • FunAudioLLM/CosyVoice2-0.5B - Provides natural, emotional, and streaming speech synthesis.

In the Deep Exploration Mode, the system employs a sophisticated chain of agents to ensure the content is factual, structured, and engaging.

graph TD
    %% Nodes
    User([👤 User Input])
    Planner[🧠 Planner Agent]
    SubPlanner[🔎 Sub Planner Agents]
    Outline[📝 Podcast Outline]
    Director[🎬 Director Agent]
    Roles[🗣️ Role Agents]
    ScreenWriter[✍️ ScreenWriter Agent]
    TTS[🌊 TTS Engine]
    Audio([🔊 Streamed Audio])
    Summary[📉 Summary Agent]
    Memory[(💾 Global Memory)]

    %% Flow
    User -->|Topic / Doc / URL| Planner
    Planner -->|Identify Intent & Distribute Tasks| SubPlanner
    
    subgraph Exploration [Parallel Exploration Phase]
        SubPlanner -->|Web Search & Research| SubPlanner
    end

    SubPlanner -->|Finalize Structure| Outline
    
    subgraph Production [Chapter Loop Production]
        Outline -->|Iterate Chapters| Director
        Director -->|Set Emotion & Direction| Roles
        Roles -->|Broadcast Information| Roles
        Roles -->|Raw Dialogue| ScreenWriter
        ScreenWriter -->|Polish & Format| TTS
        TTS -.->|Real-time Stream| Audio
    end

    %% Memory & Loop
    ScreenWriter --> Summary
    Summary -->|Update Context| Memory
    Memory -.->|Context Feedback| Director
    Memory -.->|Context Feedback| Roles

1. Planner Agent: Analyzes user intent and creates a high-level directive.
2. Sub Planner Agents: Execute parallel tasks (including Web Search) to gather information and draft a detailed Podcast Outline.
3. Director Agent: For each chapter, sets the emotional tone and guides the conversation flow.
4. Role Agents: Adopt specific personas and generate dialogue using an Information Broadcast mechanism to share knowledge.
5. ScreenWriter Agent: Aggregates the dialogue, polishes the script for natural flow, and hands it off to TTS.
6. TTS Engine: Converts the script into audio via streaming for low latency.
7. Summary Agent & Memory: Maintains the context of the conversation to ensure consistency across chapters.

• Core Framework: Setup AgentScope environment.
• Cast Modes: Support for Single, Dual, and Multi-agent conversations.
• Customization: Support for user-defined personas and specific voice timbres.
• Deep Exploration Mode: Integrate Sub-Planner parallel search and outline generation (Pending Merge).
• Input Handling:
  • User Topic Input
  • Document (PDF/Text) Parsing (Pending Merge)
  • URL Content Extraction (Pending Merge)
• Real-time Interaction: Enable users to interrupt and participate in the chat (Pending Merge).

• Python 3.9+
• CUDA-compatible GPU (for local LLM/TTS inference)

1. Clone the repository

   git clone https://github.com//h2h2h/multi_agents_podcast.git
   cd multi_agents_podcast

2. Install dependencies

   pip install -r requirements.txt

3. Model Setup

   • Download CosyVoice2-0.5B and place it in the models/tts directory.
   • Configure your LLM endpoints (or local paths for Qwen/Hunyuan) in config.yaml.

Todo

Contributions are welcome! Please feel free to submit a Pull Request.

This project is licensed under the MIT License.