wanwu/deer-flow
1
0
Fork 0
mirror of https://gitee.com/wanwujie/deer-flow synced 2026-04-24 06:34:46 +08:00
Code Issues Packages Projects Releases Wiki Activity

refactor: simplify podcast-generation to use direct JSON script input

Browse Source 
- Remove LLM script generation from Python script, model now generates
  JSON script directly (similar to image-generation skill)
- Add --transcript-file option to generate markdown transcript
- Add optional "title" field in JSON for transcript heading
- Remove dependency on OPENAI_API_KEY for podcast generation
- Update SKILL.md with new workflow and JSON format documentation

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
This commit is contained in:
hetao
2026-01-26 14:01:48 +08:00
parent 9f5658fa0e
commit dddd745b5b
3 changed files with 142 additions and 205 deletions
Download Patch File Download Diff File Expand all files Collapse all files

131
skills/public/podcast-generation/SKILL.md
View File

@@ -7,7 +7,7 @@ description: Use this skill when the user requests to generate, create, or produ
## Overview ## Overview
This skill generates high-quality podcast audio from text content using a multi-stage pipeline. The workflow includes script generation (converting input to conversational dialogue), text-to-speech synthesis, and audio mixing to produce the final podcast. This skill generates high-quality podcast audio from text content. The workflow includes creating a structured JSON script (conversational dialogue) and executing audio generation through text-to-speech synthesis.
## Core Capabilities ## Core Capabilities
@@ -24,64 +24,127 @@ This skill generates high-quality podcast audio from text content using a multi-
When a user requests podcast generation, identify: When a user requests podcast generation, identify:
- Source content: The text/article/report to convert into a podcast - Source content: The text/article/report to convert into a podcast
- Language: English or Chinese (auto-detected from content) - Language: English or Chinese (based on content)
- Output location: Where to save the generated podcast - Output location: Where to save the generated podcast
- You don't need to check the folder under `/mnt/user-data` - You don't need to check the folder under `/mnt/user-data`
### Step 2: Prepare Input Content ### Step 2: Create Structured Script JSON
The input content should be plain text or markdown. Save it to a text file in `/mnt/user-data/workspace/` with naming pattern: `{descriptive-name}-content.md` Generate a structured JSON script file in `/mnt/user-data/workspace/` with naming pattern: `{descriptive-name}-script.json`
The JSON structure:
```json
{
"locale": "en",
"lines": [
{"speaker": "male", "paragraph": "dialogue text"},
{"speaker": "female", "paragraph": "dialogue text"}
]
}
```
### Step 3: Execute Generation ### Step 3: Execute Generation
Call the Python script directly without any concerns about timeout or the need for pre-testing: Call the Python script:
```bash ```bash
python /mnt/skills/public/podcast-generation/scripts/generate.py \ python /mnt/skills/public/podcast-generation/scripts/generate.py \
--input-file /mnt/user-data/workspace/content-file.md \ --script-file /mnt/user-data/workspace/script-file.json \
--output-file /mnt/user-data/outputs/generated-podcast.mp3 \ --output-file /mnt/user-data/outputs/generated-podcast.mp3 \
--locale en --transcript-file /mnt/user-data/outputs/generated-podcast-transcript.md
``` ```
Parameters: Parameters:
- `--input-file`: Absolute path to input text/markdown file (required) - `--script-file`: Absolute path to JSON script file (required)
- `--output-file`: Absolute path to output MP3 file (required) - `--output-file`: Absolute path to output MP3 file (required)
- `--locale`: Language locale - "en" for English or "zh" for Chinese (optional, auto-detected if not specified) - `--transcript-file`: Absolute path to output transcript markdown file (optional, but recommended)
> [!IMPORTANT] > [!IMPORTANT]
> - Execute the script in one complete call. Do NOT split the workflow into separate steps (e.g., testing script generation first, then TTS). > - Execute the script in one complete call. Do NOT split the workflow into separate steps.
> - The script handles all external API calls and audio generation internally with proper timeout management. > - The script handles all TTS API calls and audio generation internally.
> - Do NOT read the Python file, just call it with the parameters. > - Do NOT read the Python file, just call it with the parameters.
> - Always include `--transcript-file` to generate a readable transcript for the user.
## Script JSON Format
The script JSON file must follow this structure:
```json
{
"title": "The History of Artificial Intelligence",
"locale": "en",
"lines": [
{"speaker": "male", "paragraph": "Hello Deer! Welcome back to another episode."},
{"speaker": "female", "paragraph": "Hey everyone! Today we have an exciting topic to discuss."},
{"speaker": "male", "paragraph": "That's right! We're going to talk about..."}
]
}
```
Fields:
- `title`: Title of the podcast episode (optional, used as heading in transcript)
- `locale`: Language code - "en" for English or "zh" for Chinese
- `lines`: Array of dialogue lines
- `speaker`: Either "male" or "female"
- `paragraph`: The dialogue text for this speaker
## Script Writing Guidelines
When creating the script JSON, follow these guidelines:
### Format Requirements
- Only two hosts: male and female, alternating naturally
- Target runtime: approximately 10 minutes of dialogue (around 40-60 lines)
- Start with the male host saying a greeting that includes "Hello Deer"
### Tone & Style
- Natural, conversational dialogue - like two friends chatting
- Use casual expressions and conversational transitions
- Avoid overly formal language or academic tone
- Include reactions, follow-up questions, and natural interjections
### Content Guidelines
- Frequent back-and-forth between hosts
- Keep sentences short and easy to follow when spoken
- Plain text only - no markdown formatting in the output
- Translate technical concepts into accessible language
- No mathematical formulas, code, or complex notation
- Make content engaging and accessible for audio-only listeners
- Exclude meta information like dates, author names, or document structure
## Podcast Generation Example ## Podcast Generation Example
User request: "Generate a podcast about the history of artificial intelligence" User request: "Generate a podcast about the history of artificial intelligence"
Step 1: Create content file `/mnt/user-data/workspace/ai-history-content.md` with the source text: Step 1: Create script file `/mnt/user-data/workspace/ai-history-script.json`:
```markdown ```json
# The History of Artificial Intelligence {
"title": "The History of Artificial Intelligence",
Artificial intelligence has a rich history spanning over seven decades... "locale": "en",
"lines": [
## Early Beginnings (1950s) {"speaker": "male", "paragraph": "Hello Deer! Welcome back to another fascinating episode. Today we're diving into something that's literally shaping our future - the history of artificial intelligence."},
The term "artificial intelligence" was coined by John McCarthy in 1956... {"speaker": "female", "paragraph": "Oh, I love this topic! You know, AI feels so modern, but it actually has roots going back over seventy years."},
{"speaker": "male", "paragraph": "Exactly! It all started back in the 1950s. The term artificial intelligence was actually coined by John McCarthy in 1956 at a famous conference at Dartmouth."},
## The First AI Winter (1970s) {"speaker": "female", "paragraph": "Wait, so they were already thinking about machines that could think back then? That's incredible!"},
After initial enthusiasm, AI research faced significant setbacks... {"speaker": "male", "paragraph": "Right? The early pioneers were so optimistic. They thought we'd have human-level AI within a generation."},
{"speaker": "female", "paragraph": "But things didn't quite work out that way, did they?"},
## Modern Era (2010s-Present) {"speaker": "male", "paragraph": "No, not at all. The 1970s brought what's called the first AI winter..."}
Deep learning revolutionized the field with breakthrough results... ]
}
``` ```
Step 2: Execute generation: Step 2: Execute generation:
```bash ```bash
python /mnt/skills/public/podcast-generation/scripts/generate.py \ python /mnt/skills/public/podcast-generation/scripts/generate.py \
--input-file /mnt/user-data/workspace/ai-history-content.md \ --script-file /mnt/user-data/workspace/ai-history-script.json \
--output-file /mnt/user-data/outputs/ai-history-podcast.mp3 \ --output-file /mnt/user-data/outputs/ai-history-podcast.mp3 \
--locale en --transcript-file /mnt/user-data/outputs/ai-history-transcript.md
``` ```
This will generate:
- `ai-history-podcast.mp3`: The audio podcast file
- `ai-history-transcript.md`: A readable markdown transcript of the podcast
## Specific Templates ## Specific Templates
Read the following template file only when matching the user request. Read the following template file only when matching the user request.
@@ -101,15 +164,14 @@ The generated podcast follows the "Hello Deer" format:
After generation: After generation:
- Podcasts are saved in `/mnt/user-data/outputs/` - Podcasts and transcripts are saved in `/mnt/user-data/outputs/`
- Share generated podcast with user using `present_files` tool - Share both the podcast MP3 and transcript MD with user using `present_files` tool
- Provide brief description of the generation result (topic, duration, hosts) - Provide brief description of the generation result (topic, duration, hosts)
- Offer to regenerate if adjustments needed - Offer to regenerate if adjustments needed
## Requirements ## Requirements
The following environment variables must be set: The following environment variables must be set:
- `OPENAI_API_KEY` or equivalent LLM API key for script generation
- `VOLCENGINE_TTS_APPID`: Volcengine TTS application ID - `VOLCENGINE_TTS_APPID`: Volcengine TTS application ID
- `VOLCENGINE_TTS_ACCESS_TOKEN`: Volcengine TTS access token - `VOLCENGINE_TTS_ACCESS_TOKEN`: Volcengine TTS access token
- `VOLCENGINE_TTS_CLUSTER`: Volcengine TTS cluster (optional, defaults to "volcano_tts") - `VOLCENGINE_TTS_CLUSTER`: Volcengine TTS cluster (optional, defaults to "volcano_tts")
@@ -117,8 +179,7 @@ The following environment variables must be set:
## Notes ## Notes
- **Always execute the full pipeline in one call** - no need to test individual steps or worry about timeouts - **Always execute the full pipeline in one call** - no need to test individual steps or worry about timeouts
- Input content language is auto-detected and matched in output - The script JSON should match the content language (en or zh)
- The script generation uses LLM to create natural conversational dialogue - Technical content should be simplified for audio accessibility in the script
- Technical content is automatically simplified for audio accessibility - Complex notations (formulas, code) should be translated to plain language in the script
- Complex notations (formulas, code) are translated to plain language
- Long content may result in longer podcasts - Long content may result in longer podcasts

212
skills/public/podcast-generation/scripts/generate.py
View File

@@ -3,7 +3,6 @@ import base64
import json import json
import logging import logging
import os import os
import re
import uuid import uuid
from typing import Literal, Optional from typing import Literal, Optional
@@ -21,7 +20,7 @@ class ScriptLine:
class Script: class Script:
def __init__(self, locale: Literal["en", "zh"] = "en", lines: list[ScriptLine] = None): def __init__(self, locale: Literal["en", "zh"] = "en", lines: Optional[list[ScriptLine]] = None):
self.locale = locale self.locale = locale
self.lines = lines or [] self.lines = lines or []
@@ -38,139 +37,6 @@ class Script:
return script return script
# Prompt template for script generation
SCRIPT_WRITER_PROMPT = """You are a skilled podcast script writer for "Hello Deer", a conversational podcast show with two hosts.
Transform the provided content into an engaging podcast script following these guidelines:
## Format Requirements
- Output as JSON with this structure: {{"locale": "en" or "zh", "lines": [{{"speaker": "male" or "female", "paragraph": "dialogue text"}}]}}
- Only two hosts: male and female, alternating naturally
- Target runtime: approximately 10 minutes of dialogue
- Start with the male host saying a greeting that includes "Hello Deer"
## Tone & Style
- Natural, conversational dialogue - like two friends chatting
- Use casual expressions and conversational transitions
- Avoid overly formal language or academic tone
- Include reactions, follow-up questions, and natural interjections
## Content Guidelines
- Frequent back-and-forth between hosts
- Keep sentences short and easy to follow when spoken
- Plain text only - no markdown formatting in the output
- Translate technical concepts into accessible language
- No mathematical formulas, code, or complex notation
- Make content engaging and accessible for audio-only listeners
- Exclude meta information like dates, author names, or document structure
## Language
- Match the locale of the input content
- Use "{locale}" for the output locale
Now transform this content into a podcast script:
{content}
"""
def extract_json_from_text(text: str) -> dict:
"""Extract JSON from text that might contain markdown code blocks or extra content."""
# Try to find JSON in markdown code blocks first
json_block_pattern = r"```(?:json)?\s*(\{[\s\S]*?\})\s*```"
match = re.search(json_block_pattern, text)
if match:
return json.loads(match.group(1))
# Try to find raw JSON object
json_pattern = r"\{[\s\S]*\}"
match = re.search(json_pattern, text)
if match:
return json.loads(match.group(0))
# Last resort: try parsing the whole text
return json.loads(text)
def generate_script(content: str, locale: str) -> Script:
"""Generate podcast script from content using LLM."""
logger.info("Generating podcast script...")
api_key = os.getenv("OPENAI_API_KEY")
base_url = os.getenv("OPENAI_BASE_URL", "https://api.openai.com/v1")
model = os.getenv("OPENAI_MODEL", "gpt-4o")
if not api_key:
raise ValueError("OPENAI_API_KEY environment variable is not set")
prompt = SCRIPT_WRITER_PROMPT.format(content=content, locale=locale)
# First try with JSON mode
try:
response = requests.post(
f"{base_url}/chat/completions",
headers={
"Authorization": f"Bearer {api_key}",
"Content-Type": "application/json",
},
json={
"model": model,
"messages": [
{"role": "system", "content": "You are a podcast script writer. Always respond with valid JSON only, no markdown formatting."},
{"role": "user", "content": prompt},
],
"response_format": {"type": "json_object"},
},
)
if response.status_code != 200:
raise Exception(f"LLM API error: {response.status_code} - {response.text}")
result = response.json()
logger.info(f"API response keys: {result.keys()}")
if "error" in result:
raise Exception(f"API error: {result['error']}")
response_content = result["choices"][0]["message"]["content"]
logger.info(f"LLM response preview: {response_content[:200]}...")
script_json = json.loads(response_content)
except (json.JSONDecodeError, KeyError) as e:
# Fallback: try without JSON mode for models that don't support it
logger.warning(f"JSON mode failed ({e}), trying without response_format...")
response = requests.post(
f"{base_url}/chat/completions",
headers={
"Authorization": f"Bearer {api_key}",
"Content-Type": "application/json",
},
json={
"model": model,
"messages": [
{"role": "system", "content": "You are a podcast script writer. Respond with valid JSON only, no markdown or extra text."},
{"role": "user", "content": prompt},
],
},
)
if response.status_code != 200:
raise Exception(f"LLM API error: {response.status_code} - {response.text}")
result = response.json()
response_content = result["choices"][0]["message"]["content"]
logger.debug(f"LLM response (fallback): {response_content[:500]}...")
script_json = extract_json_from_text(response_content)
# Validate structure
if "lines" not in script_json:
raise ValueError(f"Invalid script format: missing 'lines' key. Got keys: {list(script_json.keys())}")
script = Script.from_dict(script_json)
logger.info(f"Generated script with {len(script.lines)} lines")
return script
def text_to_speech(text: str, voice_type: str) -> Optional[bytes]: def text_to_speech(text: str, voice_type: str) -> Optional[bytes]:
"""Convert text to speech using Volcengine TTS.""" """Convert text to speech using Volcengine TTS."""
app_id = os.getenv("VOLCENGINE_TTS_APPID") app_id = os.getenv("VOLCENGINE_TTS_APPID")
@@ -264,45 +130,53 @@ def mix_audio(audio_chunks: list[bytes]) -> bytes:
return output return output
def detect_locale(content: str) -> str: def generate_markdown(script: Script, title: str = "Podcast Script") -> str:
"""Auto-detect content locale based on character analysis.""" """Generate a markdown script from the podcast script."""
chinese_chars = sum(1 for char in content if "\u4e00" <= char <= "\u9fff") lines = [f"# {title}", ""]
total_chars = len(content)
if total_chars > 0 and chinese_chars / total_chars > 0.1: for line in script.lines:
return "zh" speaker_name = "**Host (Male)**" if line.speaker == "male" else "**Host (Female)**"
return "en" lines.append(f"{speaker_name}: {line.paragraph}")
lines.append("")
return "\n".join(lines)
def generate_podcast( def generate_podcast(
input_file: str, script_file: str,
output_file: str, output_file: str,
locale: Optional[str] = None, transcript_file: Optional[str] = None,
) -> str: ) -> str:
"""Generate a podcast from input content.""" """Generate a podcast from a script JSON file."""
# Read input content # Read script JSON
with open(input_file, "r", encoding="utf-8") as f: with open(script_file, "r", encoding="utf-8") as f:
content = f.read() script_json = json.load(f)
if not content.strip(): if "lines" not in script_json:
raise ValueError("Input file is empty") raise ValueError(f"Invalid script format: missing 'lines' key. Got keys: {list(script_json.keys())}")
# Auto-detect locale if not specified script = Script.from_dict(script_json)
if not locale: logger.info(f"Loaded script with {len(script.lines)} lines")
locale = detect_locale(content)
logger.info(f"Auto-detected locale: {locale}")
# Step 1: Generate script # Generate transcript markdown if requested
script = generate_script(content, locale) if transcript_file:
title = script_json.get("title", "Podcast Script")
markdown_content = generate_markdown(script, title)
transcript_dir = os.path.dirname(transcript_file)
if transcript_dir:
os.makedirs(transcript_dir, exist_ok=True)
with open(transcript_file, "w", encoding="utf-8") as f:
f.write(markdown_content)
logger.info(f"Generated transcript to {transcript_file}")
# Step 2: Convert to audio # Convert to audio
audio_chunks = tts_node(script) audio_chunks = tts_node(script)
if not audio_chunks: if not audio_chunks:
raise Exception("Failed to generate any audio") raise Exception("Failed to generate any audio")
# Step 3: Mix audio # Mix audio
output_audio = mix_audio(audio_chunks) output_audio = mix_audio(audio_chunks)
# Save output # Save output
@@ -312,15 +186,18 @@ def generate_podcast(
with open(output_file, "wb") as f: with open(output_file, "wb") as f:
f.write(output_audio) f.write(output_audio)
return f"Successfully generated podcast to {output_file}" result = f"Successfully generated podcast to {output_file}"
if transcript_file:
result += f" and transcript to {transcript_file}"
return result
if __name__ == "__main__": if __name__ == "__main__":
parser = argparse.ArgumentParser(description="Generate podcast from text content") parser = argparse.ArgumentParser(description="Generate podcast from script JSON file")
parser.add_argument( parser.add_argument(
"--input-file", "--script-file",
required=True, required=True,
help="Absolute path to input text/markdown file", help="Absolute path to script JSON file",
) )
parser.add_argument( parser.add_argument(
"--output-file", "--output-file",
@@ -328,19 +205,18 @@ if __name__ == "__main__":
help="Output path for generated podcast MP3", help="Output path for generated podcast MP3",
) )
parser.add_argument( parser.add_argument(
"--locale", "--transcript-file",
choices=["en", "zh"], required=False,
default=None, help="Output path for transcript markdown file (optional)",
help="Language locale (auto-detected if not specified)",
) )
args = parser.parse_args() args = parser.parse_args()
try: try:
result = generate_podcast( result = generate_podcast(
args.input_file, args.script_file,
args.output_file, args.output_file,
args.locale, args.transcript_file,
) )
print(result) print(result)
except Exception as e: except Exception as e:

4
skills/public/podcast-generation/templates/tech-explainer.md
View File

@@ -49,9 +49,9 @@ This is commonly used in signup flows, admin dashboards, or when importing users
```bash ```bash
python /mnt/skills/public/podcast-generation/scripts/generate.py \ python /mnt/skills/public/podcast-generation/scripts/generate.py \
--input-file /mnt/user-data/workspace/tech-content.md \ --script-file /mnt/user-data/workspace/tech-explainer-script.json \
--output-file /mnt/user-data/outputs/tech-explainer-podcast.mp3 \ --output-file /mnt/user-data/outputs/tech-explainer-podcast.mp3 \
--locale en --transcript-file /mnt/user-data/outputs/tech-explainer-transcript.md
``` ```
## Tips for Technical Podcasts ## Tips for Technical Podcasts