Building an AI FM Radio from Scratch: A Developer's Guide to Smart Podcast System with GPT and Suno

Have you ever thought about spending a few hundred dollars on a server to have a 24/7 FM radio that automatically generates content? In the past, this required professional broadcasters, composers, and live production teams, but now the maturity of multimodal large models allows individual developers to easily build it. This article breaks down a complete AI FM radio system, from architecture design to GPT copywriting, Suno music creation, MiniMax speech synthesis, and multi-platform deployment, enabling you to quickly master the core development path.

System Architecture: Mapping the Technical Blueprint of AI FM Radio

A commercially viable AI FM radio typically consists of three layers:

• Frontend Layer: Handles user interaction, supporting websites (React/Vue), WeChat Mini Programs, cross-platform apps (Flutter), and H5 pages.
• Service Layer: Backend responsible for scheduling, task management, and user data storage (Node.js/Django/Spring Boot).
• AI Capability Layer: Integrates APIs of various large models, including copy generation (GPT/DeepSeek/GLM), speech synthesis (iFlytek Spark Voice/MiniMax Voice), music generation (Suno/Stable Audio), and image generation (Midjourney/DALL·E).

Core workflow: Customize program schedule by user → Backend triggers scheduled tasks → Call GPT to generate news/story/interview scripts → Call Suno to generate background music and mix audio → Call MiniMax voice synthesis for anchor audio → Package as streaming media output to frontend.

Key Technology Selection

• Speech Synthesis: MiniMax Voice offers multiple timbres and emotion control with low latency; iFlytek Spark Voice supports dialects and real-time streaming output.
• Copy Generation: GPT-4/DeepSeek-Chat excels in long-form text creation; GLM-130B supports Chinese style control.
• Music Generation: Suno v4 can generate full tracks based on lyrics and style; Stable Audio supports audio clip splicing.
• Multimodal Combination: Midjourney/Tongyi Wanxiang generates album covers or program illustrations; Pika/Seedance creates short video interstitials.

Core Module Implementation: Step-by-Step Guide to Building Each Feature

Module 1: Intelligent Copy Generation

The soul of a radio show is content. We achieve automated news broadcasts, book sharing, emotional talk shows, etc., through the GPT series models. Taking "AI News Flash" as an example:

• Use Python to call the OpenAI API and set a prompt template:
  prompt = "Please generate a 5-minute oral news script based on the following summary, with a lively and natural tone, including an opening and closing. Summary: {stories}"
• Split the generated text into sentences, annotate emotions (e.g., sadness, excitement), and pass them to the backend speech engine.
• To avoid repetition, you can introduce GLM's text diversity parameters or use DeepSeek's long-context capability to integrate multiple source materials.

Module 2: Multi-Timbre Speech Synthesis

Different shows require different anchor personas. MiniMax Voice provides a rich timbre library and emotion adjustment parameters:

• Register for a MiniMax developer account and obtain an API Key.
• Call the speech synthesis interface: pass in text, speed, and timbre ID (e.g., "female_calm" or "male_enthusiastic").
• For long texts, use streaming segmented synthesis—synthesize and push to the frontend simultaneously to reduce user wait time.
• Fallback: iFlytek Spark Voice supports real-time SSML (Speech Synthesis Markup Language), allowing precise control of pauses and emphasis.

Module 3: Background Music and Sound Effects Auto-Generation

Suno and Stable Audio can generate exclusive BGM based on emotion type. For example:

• In Suno, use the API to specify a style (e.g., "ambient, warm" or "electronic, upbeat") and input lyrics (can be left blank for the model to freely compose).
• Convert the generated audio to MP3 and store it in cloud storage (Alibaba Cloud OSS/MinIO).
• The backend mixes background music as an overlay during speech synthesis, adjusting the volume ratio (speech 70%, music 30%).
• Note on copyright: Use open-source models (e.g., MusicGen) or commercial APIs (Suno) to ensure compliance.

Module 4: Interactive Dialogue and Real-Time On-Demand

Allow listeners to interact with the AI anchor via voice or text, enabling "song requests" or "Q&A."

• User says "I want to hear a song by Jay Chou" → Mini-program captures voice → Call iFlytek voice recognition → Pass to GPT for intent parsing → Trigger song search (integrate with copyright library or generate similar style via Suno).
• Alternatively, use Claude's chain-of-thought capabilities to dynamically insert user comment replies during program generation, enhancing companionship.

Multi-Platform Deployment: Full Coverage from Web to Mini Program

Website Platform (H5/PC)

• Frontend uses Vue3 + HLS.js (for streaming playback); backend uses Nginx + ffmpeg for real-time transcoding.
• Push program status via WebSocket for real-time displays like "Now Playing."

WeChat Mini Program

• Use the WeChat native component live-player to play RTMP streams, or the audio component to play HLS.
• Note the network restrictions of mini programs: legally configured domain names are required, and audio duration cannot exceed 5 minutes (can be segmented).

Cross-Platform App

• Developed with Flutter, integrated with a local player package (e.g., just_audio).
• Use Firebase Cloud Messaging or a self-built push service to enable continuous background playback.

Performance Optimization and Cost Control

• Caching Strategy: Pre-generate and cache regular programs (e.g., hourly news) to CDN to reduce real-time API calls.
• Model Mixing: Use lightweight models (DeepSeek-R1) for simple content and GPT-4 for creative content to reduce costs.
• Audio/Video Compression: Use Opus encoding to provide 64kbps AAC audio, balancing quality and bandwidth.

During development, you can leverage the AI application development framework provided by Xiyue Company to accelerate integration. This framework encapsulates mainstream large model SDKs and includes a built-in task scheduling engine.

Expansion: From FM Radio to Multimodal Interactive Platform

AI FM radio is just the starting point. You can add visual elements to make the anchor "appear on screen": achieve a "visual radio" through digital human technology (e.g., MetaHuman + Sora-generated video clips). Or use Pika to create illustration carousels based on program themes. In the future, it can even combine Seedance's long video generation capabilities to create complete AI TV series.

Implementing an AI FM radio does not require a large team: one developer, three or four large model APIs, plus cloud services, is enough to run an MVP. The key is to define the content positioning clearly and reasonably orchestrate the model workflow. If you are planning a similar project, we suggest starting with a WeChat Mini Program—low user acquisition cost and quick validation of demand. Remember, technology is just a tool; what truly keeps listeners engaged is the quality and warmth of the content. Now, open your editor and start building.