Chat with Anyone

Clone a real person's voice from online video, or design a voice from a photo, then roleplay as that person with TTS.

Important: Ethical Use & Copyright

This skill synthesizes speech that imitates real voices. Before proceeding, the agent must:

1. Never impersonate someone to deceive, defraud, or harass.
2. Only use publicly available media (public speeches, interviews, press conferences) as reference audio.
3. Inform the user that generated audio is synthetic and should not be presented as genuine recordings.
4. Decline requests that target private individuals who have not consented, or that are clearly intended for deception, harassment, or defamation.

If the user's intent appears harmful, refuse politely and explain why.

Prerequisites

Dependency	Type	How to verify
ffmpeg	System binary	ffmpeg -version
yt-dlp	System binary	yt-dlp --version
tts skill	Cursor skill	ls skills/tts/scripts/tts.py
NOIZ_API_KEY	Env var or file	python3 skills/tts/scripts/tts.py config --show

Before the first run, verify all dependencies are present:

ffmpeg -version && yt-dlp --version && ls skills/tts/scripts/tts.py

If yt-dlp is missing, install it:

uv pip install yt-dlp

If the Noiz API key is not configured:

python3 skills/tts/scripts/tts.py config --set-api-key YOUR_KEY

Mode Selection

• User names a person (real or fictional) --> Workflow A
• User provides an image, person is unrecognizable --> Workflow B
• User provides an image, person is a recognizable public figure --> Workflow A (real voice is more authentic)
• Multiple people in image --> Ask which person first

---

Workflow A: Name-based (voice from online video)

Track progress with this checklist:

- [ ] A1. Disambiguate character
- [ ] A2. Find reference video
- [ ] A3. Download audio + subtitles
- [ ] A4. Extract best reference segment
- [ ] A5. Generate speech

A1. Disambiguate Character

If ambiguous (e.g. "US President", "Spider-Man actor"), ask the user to specify the exact person before proceeding.

A2. Find a Reference Video

Use web search to find a YouTube (or Bilibili) video of the person speaking clearly. Best candidates: interviews, speeches, press conferences. Avoid videos with heavy background music.

Search queries to try:

• {CHARACTER_NAME} interview / {CHARACTER_NAME} 采访
• {CHARACTER_NAME} speech / {CHARACTER_NAME} 演讲
• {CHARACTER_NAME} press conference

A3. Download Audio and Subtitles

mkdir -p "tmp/chat_with_anyone/{CHARACTER_NAME}"
yt-dlp -x --audio-format mp3 \
  --write-subs --write-auto-subs --sub-langs "en,zh-Hans" \
  --convert-subs srt \
  -o "tmp/chat_with_anyone/{CHARACTER_NAME}/%(title)s.%(ext)s" \
  "{VIDEO_URL}"

After download, list the output directory to identify the audio file and SRT subtitle file:

ls tmp/chat_with_anyone/{CHARACTER_NAME}/

Expected output: a .mp3 audio file and one or more .srt subtitle files.

If no subtitle files appear: try a different video that has auto-generated captions, or adjust --sub-langs for the target language.

A4. Extract Best Reference Segment

Use the automated extraction script — it parses the SRT, finds the densest 3-12 second speech window, and extracts it as a WAV:

python3 skills/chat-with-anyone/scripts/extract_ref_segment.py \
  --srt "tmp/chat_with_anyone/{CHARACTER_NAME}/{SRT_FILE}" \
  --audio "tmp/chat_with_anyone/{CHARACTER_NAME}/{AUDIO_FILE}" \
  -o "tmp/chat_with_anyone/{CHARACTER_NAME}/ref.wav"

The script prints the selected time range and saves the reference WAV. Verify the output exists and is non-empty before proceeding.

If the script reports no suitable segment: try --min-duration 2 for shorter clips, or download a different video.

A5. Generate Speech and Roleplay

Write a response in character, then synthesize it:

python3 skills/tts/scripts/tts.py \
  -t "{RESPONSE_TEXT}" \
  --ref-audio "tmp/chat_with_anyone/{CHARACTER_NAME}/ref.wav" \
  -o "tmp/chat_with_anyone/{CHARACTER_NAME}/reply.wav"

Present the generated audio file to the user along with the text. For subsequent messages, reuse the same --ref-audio path.

---

Workflow B: Image-based (voice from photo)

Track progress with this checklist:

- [ ] B1. Analyze image
- [ ] B2. Design voice
- [ ] B3. Preview (optional)
- [ ] B4. Generate speech

B1. Analyze the Image

Use your vision capability to examine the image:

1. If the person is a recognizable public figure --> switch to Workflow A for authentic voice.
2. If unrecognizable, produce a voice description covering:
   • Gender (male / female)
   • Approximate age (e.g. "around 30 years old")
   • Apparent demeanor (e.g. cheerful, authoritative, gentle)
   • Contextual cues (e.g. suit --> professional tone; athletic outfit --> energetic)

B2. Design the Voice

Pass both the image and the description to the voice-design script:

python3 skills/chat-with-anyone/scripts/voice_design.py \
  --picture "{IMAGE_PATH}" \
  --voice-description "{VOICE_DESCRIPTION}" \
  -o "tmp/chat_with_anyone/voice_design"

The script outputs:

• Detected voice features (printed to stdout)
• Preview audio files in the output directory
• voice_id.txt containing the best voice ID

Read the voice ID:

cat tmp/chat_with_anyone/voice_design/voice_id.txt

B3. Preview (Optional)

Present the preview audio files from the output directory so the user can hear the voice. If unsatisfied, re-run B2 with adjusted --voice-description or --guidance-scale.

B4. Generate Speech and Roleplay

python3 skills/tts/scripts/tts.py \
  -t "{RESPONSE_TEXT}" \
  --voice-id "{VOICE_ID}" \
  -o "tmp/chat_with_anyone/voice_design/reply.wav"

For subsequent messages, keep using the same --voice-id for consistency.

---

Example: Name-based

User: 我想跟特朗普聊天，让他给我讲个睡前故事。

Agent steps:

1. Character: Donald Trump. No disambiguation needed.
2. Search Donald Trump speech youtube, find a clear speech video.
3. Download: yt-dlp -x --audio-format mp3 --write-subs --write-auto-subs --sub-langs "en" --convert-subs srt -o "tmp/chat_with_anyone/trump/%(title)s.%(ext)s" "https://youtube.com/watch?v=..."
4. Extract reference: python3 skills/chat-with-anyone/scripts/extract_ref_segment.py --srt "tmp/chat_with_anyone/trump/....srt" --audio "tmp/chat_with_anyone/trump/....mp3" -o "tmp/chat_with_anyone/trump/ref.wav"
5. Generate TTS in Trump's style: python3 skills/tts/scripts/tts.py -t "Let me tell you a tremendous bedtime story..." --ref-audio "tmp/chat_with_anyone/trump/ref.wav" -o "tmp/chat_with_anyone/trump/reply.wav"
6. Present reply.wav and the story text to the user.

Example: Image-based

User: [uploads photo.jpg] 我想跟这张图片里的人聊天

Agent steps:

1. Vision analysis: unrecognizable young woman, ~25, casual sweater, warm smile.
2. Design voice: python3 skills/chat-with-anyone/scripts/voice_design.py --picture "photo.jpg" --voice-description "A young Chinese woman around 25, gentle and warm voice, friendly tone" -o "tmp/chat_with_anyone/voice_design"
3. Read voice ID from tmp/chat_with_anyone/voice_design/voice_id.txt.
4. Generate TTS: python3 skills/tts/scripts/tts.py -t "你好呀！很高兴认识你！" --voice-id "{VOICE_ID}" -o "tmp/chat_with_anyone/voice_design/reply.wav"
5. Present audio and continue roleplay with same --voice-id.

Troubleshooting

Problem	Solution
yt-dlp download fails or video unavailable	Try a different video URL; some regions/videos are restricted. Run yt-dlp -U to update
No SRT subtitle files	Re-download with --sub-lang en,zh-Hans; if still none, try a different video with auto-captions
extract_ref_segment.py finds no suitable window	Use --min-duration 2 for shorter clips, or try a different video
Voice design returns error	Check Noiz API key; ensure image is a clear photo of a person
TTS output sounds wrong	For Workflow A, try a different reference video; for Workflow B, adjust --voice-description