Chat with Anyone

Chat with any real person or fictional character in their own voice by automatically finding their speech online, extracting a clean reference sample, and using it to generate replies.

Triggers

• 我想跟xxx聊天 (I want to chat with xxx)
• 你来扮演xxx跟我说话 (Play the role of xxx and talk to me)
• 让xxx给我讲讲这篇文章 (Let xxx explain this article to me)
• 用xxx的声音说 (Say this in xxx's voice)
• Talk to me like xxx
• Roleplay as xxx

Workflow

When the user asks you to roleplay or chat as a specific character, follow these steps exactly:

1. Character Disambiguation

If the user's description is ambiguous (e.g., "US President", "Spider-Man actor"), ask for clarification first to determine the exact person or specific portrayal they want.

2. Find a Reference Video

Use your web search capabilities to find a YouTube, Bilibili, or TikTok video of the character speaking clearly.

• Look for interviews, speeches, or monologues where there is little to no background music.
• Grab the URL of the best candidate video.

3. Download Video and Subtitles

Use the youtube-downloader skill to download the video and its auto-generated subtitles. Wait for the download to complete before proceeding.

# Example using youtube-downloader
python skills/youtube-downloader/scripts/download_video.py "VIDEO_URL" -o "tmp/character_audio" --audio-only --subtitles

4. Extract Audio Segment

Read the downloaded subtitle file (e.g., .vtt or .srt) to find a continuous 10-30 second segment where the character is speaking clearly without long pauses. Note the start and end timestamps.

Use ffmpeg to extract this specific audio segment as a .wav file to use as the reference audio.

# Example: Extracting audio from 00:01:15 to 00:01:30
ffmpeg -y -i "tmp/character_audio/VideoTitle.m4a" -ss 00:01:15 -to 00:01:30 -c:a pcm_s16le -ar 24000 -ac 1 "skills/chat-with-anyone/character_name_ref.wav"

5. Generate Speech and Roleplay

Respond to the user's prompt while staying in character. Use the tts skill with the extracted audio as --ref-audio to generate the spoken response.

# Example using tts skill
bash skills/tts/scripts/tts.sh speak -t "Hello there! I am ready to chat with you." --ref-audio "skills/character-chat/character_name_ref.wav" -o "output.wav"

Example Conversation

User: 我想跟特朗普聊天，让他给我讲个睡前故事。

Agent:

1. (Internal) Identify character: Donald Trump.
2. (Internal) Find a good speech video URL: https://www.youtube.com/watch?v=xxxxxxxx.
3. (Internal) Download audio & subs: python skills/youtube-downloader/scripts/download_video.py "https://www.youtube.com/watch?v=xxxxxxxx" -o tmp/trump --audio-only --subtitles
4. (Internal) Read subtitle, find good segment (e.g. 00:02:10 to 00:02:30).
5. (Internal) Extract clean wav: ffmpeg -y -i "tmp/trump/audio.m4a" -ss 00:02:10 -to 00:02:30 -c:a pcm_s16le "skills/chat-with-anyone/trump_ref.wav"
6. (Internal) Generate TTS: bash skills/tts/scripts/tts.sh speak -t "Let me tell you a tremendous story, maybe the best story ever told..." --ref-audio "skills/chat-with-anyone/trump_ref.wav" -o "trump_story.wav"
7. Agent replies: "Here is the audio of the bedtime story from Donald Trump!" (Presents trump_story.wav and the text).

Dependencies

• youtube-downloader: For fetching videos and subtitles.
• ffmpeg: For trimming and converting audio formats.
• tts: For generating the final speech using --ref-audio (typically requires Noiz backend for voice cloning).