Talking Circle

Create animated talking-circle videos (Telegram-style round video messages) from avatar frame images and audio. Supports audio-to-video and text-to-video via ElevenLabs or SaluteSpeech (Sber) TTS. Use when the user wants to generate lip-synced circular avatar animations, talking circles, or round video messages.

Create animated circular avatar videos with lip-sync and blink animations. Takes 4 avatar frame images (neutral, slight open, wide open, eyes closed) and produces a round video with audio-driven mouth movement.

Prerequisites

• python3 (3.9+)
• ffmpeg installed and on PATH
• Optional: ELEVENLABS_API_KEY environment variable (for ElevenLabs text-to-video mode)
• Optional: SALUTE_SPEECH_AUTH environment variable (for SaluteSpeech text-to-video mode)

Setup

Dependencies are auto-installed into a temporary venv on first run. To install manually:

bash

pip install -r requirements.txt

Mode 1: Audio to Video

Convert existing audio + frame images into an animated talking circle video.

bash

python3 scripts/make_talking_circle_video.py \
  --neutral frames/neutral.png \
  --slight frames/mouth-slight-open.png \
  --wide frames/mouth-wide-open.png \
  --blink frames/eyes-closed.png \
  --audio speech.mp3 \
  --out /tmp/talking-circle.mp4

Mode 2: Text to Video

Generate speech from text via ElevenLabs TTS, then create the animated video.

Requires ELEVENLABS_API_KEY set in environment or passed via --api-key.

bash

python3 scripts/make_text_to_video.py \
  --text "Hello, this is a talking circle demo!" \
  --voice-id pNInz6obpgDQGcFmaJgB \
  --neutral frames/neutral.png \
  --slight frames/mouth-slight-open.png \
  --wide frames/mouth-wide-open.png \
  --blink frames/eyes-closed.png \
  --out /tmp/talking-circle.mp4

Mode 3: Text to Video via SaluteSpeech (Sber)

Generate speech from text via SaluteSpeech TTS (Sber), then create the animated video.

Requires SALUTE_SPEECH_AUTH set in environment or passed via --auth-key. This is a Base64-encoded client_id:client_secret from your SaluteSpeech project.

bash

python3 scripts/make_salute_text_to_video.py \
  --text "Привет, это демонстрация talking circle!" \
  --voice Bys_24000 \
  --neutral frames/neutral.png \
  --slight frames/mouth-slight-open.png \
  --wide frames/mouth-wide-open.png \
  --blink frames/eyes-closed.png \
  --out /tmp/talking-circle.mp4

SaluteSpeech voices

Voice	Name	Language
Nec_24000	Natalia (female)	ru-RU
Bys_24000	Boris (male)	ru-RU
May_24000	Martha (female)	ru-RU
Tur_24000	Taras (male)	ru-RU
Ost_24000	Alexandra (female)	ru-RU
Pon_24000	Sergey (male)	ru-RU
Kin_24000	Kira (female)	en-US

Voice Presets

ElevenLabs preset (Sbercat — male)

Parameter	Value
--voice-id	pNInz6obpgDQGcFmaJgB
--model-id	eleven_multilingual_v2
--stability	0.15
--similarity-boost	0.70
--style	0.38
--speed	1.20

SaluteSpeech preset (Boris — male, Russian)

Parameter	Value
--voice	Bys_24000
--audio-format	wav16
--scope	SALUTE_SPEECH_PERS

How to get API keys

ElevenLabs:

1. Go to ElevenLabs Voice Library.
2. Pick or clone a voice, copy the voice ID.
3. Set ELEVENLABS_API_KEY environment variable.

SaluteSpeech (Sber):

1. Register at developers.sber.ru.
2. Create a project, get client_id and client_secret.
3. Encode client_id:client_secret in Base64.
4. Set SALUTE_SPEECH_AUTH environment variable with the Base64 string.

Alternative TTS engines

The skill also supports any TTS that can produce an audio file. Use Mode 1 (audio-to-video) with audio from any source:

• OpenAI TTS (openai.audio.speech.create) — generate speech, save to MP3, pass via --audio
• Local TTS (Coqui, Piper, Silero, etc.) — run locally, save WAV/MP3, pass via --audio
• Google Cloud TTS, Amazon Polly, Azure TTS — any cloud provider works

bash

# Example: generate audio with any TTS, then animate
python3 scripts/make_talking_circle_video.py \
  --neutral frames/neutral.png \
  --slight frames/mouth-slight-open.png \
  --wide frames/mouth-wide-open.png \
  --blink frames/eyes-closed.png \
  --audio /path/to/speech-from-any-tts.mp3 \
  --out /tmp/talking-circle.mp4

Tell the user: if they don't have an ElevenLabs or SaluteSpeech API key, they can use any other TTS engine — just generate the audio file and pass it to Mode 1. No API key needed for audio-to-video mode.

Frame Image Requirements

You need 4 PNG images of your avatar, all the same resolution (recommended 2048x2048), square aspect ratio:

Frame	Description
neutral	Mouth closed, eyes open
slight	Mouth slightly open, eyes open
wide	Mouth wide open, eyes open
blink	Mouth closed, eyes closed

Critical rules

• All 4 frames must have identical resolution, art style, colors, and character positioning.
• Only the mouth and eyes should change between frames — head, body, background must stay the same.
• Do not mix frames from different generation sessions or different styles.

Generating Frames with Image AI

If the user does not have ready-made frames, generate them using an image generation API (DALL-E, Midjourney, Flux, etc.). Follow this workflow:

Step 1: Generate the neutral frame

Generate a shoulder-up portrait of the character. This is the base frame — all other frames must match it exactly.

Example prompt:

text

Shoulder-up portrait of [CHARACTER DESCRIPTION]. Square composition, clean background,
mouth closed, eyes open, looking at camera. High detail, consistent lighting.

Step 2: Generate remaining 3 frames as edits of neutral

Use image editing / inpainting on the neutral frame to produce the other states. Only modify the mouth and eyes region — everything else must remain pixel-identical.

Frame	What to change	Edit prompt example
slight	Mouth slightly open	"Mouth slightly open, teeth barely visible, same expression"
wide	Mouth wide open	"Mouth wide open as if saying 'ah', same expression"
blink	Eyes closed	"Eyes gently closed, mouth closed, same expression"

Step 3: Verify consistency

Before using the frames:

1. Check that all 4 images have the same resolution.
2. Overlay them to verify the head/body position hasn't shifted.
3. If any frame drifts, regenerate it from the neutral base.

Examples

See examples/sbercat/ for a reference character with ready-to-use frame set:

• Sbercat — 3D-rendered anthropomorphic cat, lavender-blue fur, green eyes, pink nose, green hoodie:
  • examples/sbercat/reference.png — character reference
  • examples/sbercat/neutral.png — mouth closed, eyes open
  • examples/sbercat/slight.png — mouth slightly open
  • examples/sbercat/wide.png — mouth wide open
  • examples/sbercat/blink.png — eyes closed
  • examples/sbercat/example.mp4 — finished talking-circle video

To test with the included Sbercat frames:

bash

python3 scripts/make_talking_circle_video.py \
  --neutral examples/sbercat/neutral.png \
  --slight examples/sbercat/slight.png \
  --wide examples/sbercat/wide.png \
  --blink examples/sbercat/blink.png \
  --audio your-audio.mp3 \
  --out /tmp/talking-circle.mp4

Post-installation message

After installing this skill, inform the user:

Skill "talking-circle" installed.

To create talking-circle videos, I need a character avatar. Please send me a reference image of your character on a white/clean background — a shoulder-up portrait works best (square aspect ratio, high resolution).

From this reference I will generate 4 frame images (mouth states + blink) and you'll be ready to create animated video circles.

For speech, I can use ElevenLabs TTS (requires ELEVENLABS_API_KEY), SaluteSpeech from Sber (requires SALUTE_SPEECH_AUTH), or you can provide your own audio file. Any TTS engine works — OpenAI TTS, Whisper, Coqui, Piper, Google TTS, etc.

First use: generating frame images

IMPORTANT — before the skill can create videos, the 4 frame images must exist. If the user does not already have frames, you MUST generate them first.

Workflow for the assistant

1. Ask the user for a reference image of their character (or use a character description). The image should be a shoulder-up portrait on a white or clean background, square aspect ratio.
2. Generate the neutral frame using image generation (DALL-E, Flux, Midjourney, etc.):
   • Prompt: "Shoulder-up portrait of [CHARACTER], white background, mouth closed, eyes open, looking at camera, square composition, high detail"
   • Save as neutral.png.
3. Generate the 3 remaining frames via inpainting/editing of the neutral frame. Only modify the mouth/eyes region — everything else must remain pixel-identical:
   • slight.png — edit mouth region: "Mouth slightly open, teeth barely visible"
   • wide.png — edit mouth region: "Mouth wide open as if saying 'ah'"
   • blink.png — edit eyes region: "Eyes gently closed, mouth closed"
4. Verify consistency: all 4 images must have the same resolution, identical head/body position, and only differ in mouth/eyes.
5. Save the frames to a persistent location (e.g. the skill's working directory or a user-specified path). These frames are reused for every future video.
6. Confirm to the user that frames are ready and the skill is operational.

Do not skip this step. Without the 4 frame images, the video scripts will fail.

Guardrails

• Before running any script, verify that python3 (3.9+) and ffmpeg are on PATH. If missing, instruct the user to install them.
• Never delete or overwrite the user's original frame images.
• Do not use this skill for full-motion video editing, face tracking, or real-time lipsync — it only works with 4 static frame images.
• If ElevenLabs or SaluteSpeech API returns an error (401, 429, etc.), explain the error clearly to the user instead of retrying silently.

Failure handling

• If ffmpeg is not found: tell the user to install it (brew install ffmpeg on macOS, apt install ffmpeg on Linux).
• If ELEVENLABS_API_KEY is missing and the user wants text-to-video: suggest SaluteSpeech (Mode 3) or Mode 1 with audio from another TTS.
• If SALUTE_SPEECH_AUTH is missing and the user wants SaluteSpeech: explain how to register at developers.sber.ru and get credentials.
• If frame images have different resolutions: warn the user and ask them to fix the frames before proceeding.
• If the output video is empty or zero bytes: show the ffmpeg error log and suggest checking input files.

Parameters Reference

Video output

Parameter	Default	Description
--size	720	Output video size in pixels
--diameter	640	Circle diameter within the video
--fps	30	Frames per second

Blink timing

Parameter	Default	Description
--blink-start	1.1	Seconds before first blink
--blink-every	3.8	Seconds between blinks
--blink-duration-frames	4	Number of frames per blink

Amplitude thresholds (audio-to-video)

Parameter	Default	Description
--amp-low	1200	RMS below this = neutral (closed mouth)
--amp-high	2600	RMS above this = wide open mouth

ElevenLabs TTS settings (make_text_to_video.py)

Parameter	Default	Description
--voice-id	(required)	ElevenLabs voice ID
--model-id	eleven_multilingual_v2	ElevenLabs model
--stability	0.50	Voice stability
--similarity-boost	0.75	Voice similarity boost
--style	0.00	Style exaggeration
--speed	1.00	Speech speed

SaluteSpeech TTS settings (make_salute_text_to_video.py)

Parameter	Default	Description
--voice	Bys_24000	SaluteSpeech voice (see voices table above)
--audio-format	wav16	Audio format: opus, wav16, pcm16
--scope	SALUTE_SPEECH_PERS	OAuth scope (PERS for personal, CORP for corporate)
--auth-key	$SALUTE_SPEECH_AUTH	Base64-encoded client_id:client_secret