podcast generate

Podcast Generate Skill (TypeScript version)

Automatically generate podcast scripts and audio based on user-provided materials or web search results.

This Skill is suitable for:

Quick comprehension and podcastization of long-form content
Audio presentation of knowledge-based content
In-depth interpretation and discussion of trending topics
Searching real-time information and podcast production

Capability Description

What this Skill can do

Generate from file: Receive a piece of material (txt/md/docx/pdf and other text formats), generate a dialogue podcast script and audio
Generate via web search: Based on a topic specified by the user, search the web for the latest information, and generate a podcast script and audio
Automatically control duration, automatically adjusted according to content length (3-20 minutes)
Generate a podcast script in Markdown format (can be manually edited)
Use z-ai TTS to synthesize high-quality audio and concatenate it into the final podcast

What this Skill currently does not do

Does not generate mp3 / subtitles / timestamps
Does not support podcast roles of three or more people
Does not add background music or sound effects

Files and Responsibilities Description

This Skill consists of the following files:

generate.ts Unified entry point (supports file mode and search mode)
- File mode: Read the text file uploaded by the user → generate podcast
- Search mode: Call the web-search skill to obtain materials → generate podcast
- Use z-ai-web-dev-sdk for LLM script generation
- Use z-ai-web-dev-sdk for TTS audio generation
- Automatically concatenate audio segments
- Only output the final file
readme.md Usage documentation
SKILL.md The current file, describing the Skill's capabilities, boundaries, and usage conventions
package.json Node.js project configuration and dependencies
tsconfig.json TypeScript compilation configuration

Input and Output Conventions

Input (choose one of two)

Method 1: File upload

One material file (txt / md / docx / pdf and other text formats)
Material length is unlimited; the Skill will automatically compress it to an appropriate length

Method 2: Web search

The user specifies a search topic
Automatically call the web-search skill to obtain related content
Integrate multiple search results as material sources

Output (only 2 files are output)

podcast_script.md Podcast script (Markdown format, can be manually edited)
podcast.wav The final concatenated podcast audio

No intermediate files are output(such as segments.jsonl, meta.json, etc.)

How to Run

Dependency Environment

Node.js 18+
z-ai-web-dev-sdk (already installed)
web-search skill (used for web search mode)

Not required z-ai CLI

Install Dependencies

npm install

Usage Examples

Generate podcast from file

npm run generate -- --input=test_data/material.txt --out_dir=out

Generate podcast via web search

# 根据主题搜索并生成播客
npm run generate -- --topic="最新AI技术突破" --out_dir=out

# 指定搜索主题和时长
npm run generate -- --topic="量子计算应用场景" --out_dir=out --duration=8

# 搜索并生成单人播客
npm run generate -- --topic="气候变化影响" --out_dir=out --mode=single-male

Parameter Description

Parameter	Description	Default Value
`--input`	Input material file path (choose one of two with --topic)	-
`--topic`	Search topic keyword (choose one of two with --input)	-
`--out_dir`	Output directory (required)	-
`--mode`	Podcast mode: dual / single-male / single-female	dual
`--duration`	Manually specify minutes (3-20); 0 means automatic	0
`--host_name`	Host/Anchor name	Xiaopu
`--guest_name`	Guest name	Chuichui
`--voice_host`	Host voice	xiaochen
`--voice_guest`	Guest voice	chuichui
`--speed`	Speech rate (0.5-2.0)	1.0
`--pause_ms`	Pause between segments in milliseconds	200

Available voices

Voice	Characteristics
xiaochen	Calm and professional
chuichui	Lively and cute
tongtong	Warm and friendly
jam	British gentleman
kazi	Clear and standard
douji	Natural and fluent
luodo	Highly expressive

Technical architecture

generate.ts (unified entry point)

File mode: read user-uploaded file → generate podcast
Search mode: call web-search skill → obtain materials → generate podcast
LLM: use z-ai-web-dev-sdk (chat.completions.create)
TTS: use z-ai-web-dev-sdk (audio.tts.create)
Not required z-ai CLI
Automatically concatenate audio segments
Only output the final file; intermediate files are automatically cleaned up

LLM calls

System prompt: podcast script writer role
User prompt: includes materials + hard constraints + breathing-feel requirements
Output validation: word count, structure, role tags
Automatic retry: up to 3 times

TTS calls

Use zai.audio.tts.create()
Supports custom voices and speech rates
Automatically concatenate multiple wav segments
Temporary files are automatically cleaned up

Output example

podcast_script.md (excerpt)

**小谱**：大家好，欢迎收听今天的播客。今天我们来聊一个有趣的话题……

**锤锤**：是啊，这个话题真的很有意思。我最近也在关注……

**小谱**：说到这里，我想给大家举个例子……

License

MIT