Facial and lipsync skill for AI agents

facial-and-lipsync is a skill from the Summer skill library: a guide your AI agent reads when the task matches, so it builds the good version on the first try. Make a character’s mouth move with a voice-over: phonemes, viseme blendshapes, facial expressions.


Status	Stable
Domains	Animation, 3D models
Version	`1.0.2`
Engine version	`>=4.6`
Recommended set	No, installed by `--all` or by name

When your agent uses it

The user has (or wants) a voice-over and needs the character’s mouth to move with the words
Adding emotional facial expressions (smile, frown, surprise)

Install

summer setup installs every skill for your agent. To add only this one:

npx -y summer-engine@latest skills install facial-and-lipsync --agent claude-code

Swap claude-code for your agent, such as codex or cursor. See Skills for every option.

The skill

Half generative, half authored. The generative half: an audio file (TTS via assets.generate_audio with capability: "text_to_speech") goes through a phoneme-extraction tool (Summer does NOT wrap one — you run it externally), out comes a viseme timeline — a list of { phoneme, start_time, duration } triples. The authored half: the character’s face mesh must have BlendShapes named for the standard viseme set, or the timeline has nothing to drive. Without both halves, you get a talking robot. The pipeline:

Audio in — .wav/.mp3 from assets.generate_audio({ capability: "text_to_speech", ... }) or imported VO.
Phoneme extraction — done OUTSIDE Summer. Recommended: Rhubarb Lip Sync (open source, widely used for game lipsync; outputs JSON with mouth-shape cues over time). Alternatives: gentle, allosaurus, or a Whisper-phoneme model on Replicate / Hugging Face. There is no summer_* MCP for this in the current engine.
Viseme mapping — phonemes / Rhubarb cues → Oculus / Apple ARKit viseme set (15 standard shapes covers English).
BlendShape driver — at runtime, lerp the mesh’s BlendShape weights along the timeline, synced to audio playback.
Optional emotional layer — separate BlendShape track for smile, brow_raise, eye_squint etc., authored in code or at clip-edit time.

When to use this skill

NPC has voice lines and the mouth doesn’t move.
Cinematic with VO that needs lipsync.
“Make him sing the lyrics.”
Animating expressions on a face that has BlendShapes. Never assume it does — step 1 below is the check, and generated characters frequently come back with zero.
Reactive face: smile when player gives gold, scowl when attacked.

When NOT to use this skill

Character has no face / no BlendShapes (helmeted soldier, robot, mascot). Hard-skip to body language via generate-motion.
The dialogue is written but not yet voiced — generate audio first via assets.generate_audio({ capability: "text_to_speech", ... }). Lipsync without audio has nothing to sync to.
Pre-rendered cinematic from external DCC. Lipsync is in the rendered video, not the engine.

Steps

1. Confirm the head has BlendShapes

summer_inspect_node "./World/NPC/Head"     # or wherever the head MeshInstance3D lives

Look for mesh.blend_shape_count > 0 and a list of names. An ARKit-52 head has names like jawOpen, mouthClose, mouthFunnel, mouthPucker, mouthLeft, mouthRight, mouthSmile_L, mouthSmile_R, mouthFrown_L, mouthFrown_R, browInnerUp, browOuterUp_L, browDown_L, eyeBlink_L, eyeWide_L, eyeSquint_L (mirrored on R). If the head has zero BlendShapes, stop. Tell the user: “This head mesh has no BlendShapes — facial animation needs shape keys on the mesh. Options: use an emotional body-language overlay instead, author shape keys in Blender and re-import, or hand the character to a 3D artist.” Don’t proceed. Do not promise a face_blendshapes generation option. assets.generate_3d has no such field and nothing in the Summer generation pipeline handles one; a field the operation does not list does nothing. Regenerating will not add BlendShapes. Shape keys come from the mesh author (Blender / an ARKit-52 source mesh), not from a rig pass.

2. Get the audio

If the user has VO already, use it. If not, generate via TTS. On the ElevenLabs models the voiceId must be one of Summer’s game-cast voices (the voice-line skill lists them with their ids); any other id is refused. There is no “voice name” string; it’s always an ID like IKne3meq5aSn9XLyUdCD (Charlie, the hero).

assets.generate_audio({
  capability: "text_to_speech",
  text: "Welcome to the village, traveler.",
  voiceId: "<cast voice id>"
})
// answers at once with asset: { assetId, fileUrl, durationSeconds }

If the user hasn’t picked a voice yet, stop and ask which cast voice fits the character (see voice-line).

3. Extract phonemes (the generative step — runs OUTSIDE Summer)

Summer does not wrap a phoneme-extraction MCP tool. Run it externally, then import the result. Recommended: Rhubarb Lip Sync — an open-source CLI built for game-dev lipsync. Mouth shapes A–H map cleanly to viseme shapes. Install once:

# macOS
brew install rhubarb-lip-sync
# Windows / Linux: download release from https://github.com/DanielSWolf/rhubarb-lip-sync

Run per VO line:

rhubarb -f json -o welcome_traveler.json welcome_traveler.wav

Output JSON shape:

{
  "metadata": { "duration": 2.34 },
  "mouthCues": [
    { "start": 0.00, "end": 0.08, "value": "B" },
    { "start": 0.08, "end": 0.18, "value": "C" },
    { "start": 0.18, "end": 0.25, "value": "D" },
    { "start": 0.25, "end": 0.30, "value": "B" },
    { "start": 0.30, "end": 0.42, "value": "A" },
    { "start": 0.42, "end": 0.50, "value": "B" },
    { "start": 2.30, "end": 2.34, "value": "X" }
  ]
}

Rhubarb mouth-shape → ARKit viseme cheat sheet (apply during the bake step):

Rhubarb	ARKit viseme	Notes
A	`viseme_PP`	closed (P, B, M)
B	`viseme_kk`	slightly open, neutral
C	`viseme_E`	open, lips spread (EH, IH)
D	`viseme_aa`	wide open (AA, AH)
E	`viseme_O`	rounded (OW, ER)
F	`viseme_U`	small rounded (UW, OO)
G	`viseme_FF`	F, V (lower lip + teeth)
H	`viseme_RR`	L, R (tongue raised)
X	`viseme_sil`	silence / closed neutral

Manual fallback (no Rhubarb available): for short lines you can hand-type a mouthCues array by listening to the clip and tagging vowel/consonant boundaries. Painful past ~3s of audio; use Rhubarb for anything longer. Cloud fallback: a Whisper-phoneme model on Replicate or a Hugging Face inference endpoint will emit ARPAbet phonemes with timestamps. Then map ARPAbet → ARKit visemes via the table in the Reference card section. More accurate than Rhubarb on noisy audio; slower and costs cents per minute. Rhubarb is free and takes ~5s of local CPU per 30s clip. The cloud fallback is billed per minute by its provider and takes ~5s for a 30s clip.

4. Persist the viseme track as an AnimationLibrary entry

Convert Rhubarb’s mouthCues into a Summer Engine Animation resource: one track per BlendShape with keyframes at each viseme transition. This makes lip sync replayable through the same AnimationPlayer/AnimationTree as body motion.

# scripts/lipsync_baker.gd — run once per VO line at edit time
const RHUBARB_TO_VISEME := {
    "A": "viseme_PP",  "B": "viseme_kk", "C": "viseme_E",
    "D": "viseme_aa",  "E": "viseme_O",  "F": "viseme_U",
    "G": "viseme_FF",  "H": "viseme_RR", "X": "viseme_sil",
}

static func bake_from_rhubarb(rhubarb_json_path: String, head_path: NodePath) -> Animation:
    var f := FileAccess.open(rhubarb_json_path, FileAccess.READ)
    var data: Dictionary = JSON.parse_string(f.get_as_text())
    var cues: Array = data["mouthCues"]
    var anim := Animation.new()
    anim.length = float(data["metadata"]["duration"])
    var visemes := ["viseme_aa", "viseme_E", "viseme_I", "viseme_O", "viseme_U",
                    "viseme_PP", "viseme_FF", "viseme_TH", "viseme_DD", "viseme_kk",
                    "viseme_CH", "viseme_SS", "viseme_nn", "viseme_RR", "viseme_sil"]
    var tracks := {}
    for v in visemes:
        var idx := anim.add_track(Animation.TYPE_BLEND_SHAPE)
        anim.track_set_path(idx, NodePath(str(head_path) + ":" + v))
        tracks[v] = idx
    # For each cue, set the active viseme to 1.0 and others to 0.0 at cue.start
    for cue in cues:
        var active_viseme: String = RHUBARB_TO_VISEME.get(cue["value"], "viseme_sil")
        for v in visemes:
            var weight: float = 1.0 if v == active_viseme else 0.0
            anim.track_insert_key(tracks[v], float(cue["start"]), weight)
    return anim

Bake once, save into the character’s AnimationLibrary as dialogue_<line_id>, and play via the AnimationTree. This is a plain Animation resource authored in code — no editor required. Animation.new() + add_track() + track_set_path() + track_insert_key() + ResourceSaver.save() works from a headless -s script and the resource reloads with every track and track type intact (verified on the shipped 4.6.1 binary with value, blend-shape and method tracks in one resource). Any advice that says animation tracks or keyframes can only be authored in the editor GUI is wrong. The bake function also runs unchanged inside a summer_run_script body against the live editor — often more convenient than the headless lane because the result lands in the open scene’s library and the run is transactional. For QUICK facial keys that don’t need a viseme timeline (a roar, a wince, an eyebrow raise), animation-tier engines collapse the whole bake into ctx.animate with a "blend_shapes/<name>" property path (value tracks, verified working) — one call per shape, same anim_name appends tracks to one clip. Recipe in character-animation-wiring. The Rhubarb bake loop above stays the right tool for full audio-synced viseme timelines — 15 tracks with per-cue keys is exactly what the loop is for. If you used the Whisper/ARPAbet cloud fallback instead of Rhubarb, swap bake_from_rhubarb for a variant that consumes { phoneme, start, duration } triples and applies the ARPAbet → viseme table from the Reference card.

5. Wire into the AnimationTree

Add a OneShot node Lipsync that overlays the viseme animation as an additive layer over the base face. Fire from the dialogue system:

@onready var tree: AnimationTree = $AnimationTree
@onready var audio: AudioStreamPlayer3D = $VoicePlayer

func say(line_id: String) -> void:
    var clip_id := "dialogue_" + line_id
    tree.set("parameters/Lipsync/animation", clip_id)
    audio.stream = load("res://audio/" + line_id + ".ogg")
    audio.play()
    tree.set("parameters/Lipsync/request", AnimationNodeOneShot.ONE_SHOT_REQUEST_FIRE)

Sync is preserved as long as both fire on the same frame. ~16ms drift is the threshold of perception; AnimationTree + AudioStreamPlayer3D are both sample-accurate, so drift only happens if the engine hitches mid-line.

6. Add the emotional layer (authored, not generative)

A second OneShot or persistent additive track for expressions. Key the relevant BlendShapes (mouthSmile_L, mouthSmile_R, browInnerUp, etc.) at design time:

func smile(intensity: float) -> void:
    var head: MeshInstance3D = $Head
    head.set_blend_shape_value(head.find_blend_shape_by_name("mouthSmile_L"), intensity)
    head.set_blend_shape_value(head.find_blend_shape_by_name("mouthSmile_R"), intensity)

func surprise(intensity: float) -> void:
    var head: MeshInstance3D = $Head
    head.set_blend_shape_value(head.find_blend_shape_by_name("browInnerUp"), intensity)
    head.set_blend_shape_value(head.find_blend_shape_by_name("browOuterUp_L"), intensity * 0.7)
    head.set_blend_shape_value(head.find_blend_shape_by_name("browOuterUp_R"), intensity * 0.7)
    head.set_blend_shape_value(head.find_blend_shape_by_name("eyeWide_L"), intensity)
    head.set_blend_shape_value(head.find_blend_shape_by_name("eyeWide_R"), intensity)
    head.set_blend_shape_value(head.find_blend_shape_by_name("jawOpen"), intensity * 0.3)

Driven from gameplay events; orthogonal to the lipsync layer.

Confirmation gates

Before extracting: show audio length, est. cost, est. wait. Wait for OK.
Before baking the Animation resource: confirm the head’s BlendShape names match the standard set. If they’re custom-named, ask for the mapping.
Before saving the scene: confirm the AnimationTree changes (added Lipsync OneShot, library entry).

Reference card

ARKit-52 viseme subset (the 15 that matter for English lipsync)

Viseme	Triggered by phonemes (ARPAbet)	Mouth shape
`viseme_sil`	silence	closed neutral
`viseme_PP`	P, B, M	lips pressed
`viseme_FF`	F, V	lower lip + upper teeth
`viseme_TH`	TH, DH	tongue tip + teeth
`viseme_DD`	T, D, N, S, Z	tongue + alveolar
`viseme_kk`	K, G, NG	back-tongue, mouth slightly open
`viseme_CH`	CH, JH, SH, ZH	lips rounded forward
`viseme_SS`	S, Z (sometimes split from DD)	teeth nearly closed
`viseme_nn`	N, L	tongue-tip + open mouth
`viseme_RR`	R, ER	mouth slightly rounded
`viseme_aa`	AA, AH, AE	wide open
`viseme_E`	EH, EY, IH	mid open + lips spread
`viseme_I`	IY, IH	narrow + spread
`viseme_O`	OW, AO	rounded
`viseme_U`	UW, UH	rounded + small

Phoneme → viseme mapping (for hand-rolled extraction)

If the user has a CMUDict-style phoneme list, map it with the ARPAbet column of the table above.

Pitfalls

Mouth never closes between words. No viseme_sil keyframes at silence intervals. The bake step must scan the audio for silence (RMS below threshold for 100ms+) and insert sil keys, OR the phoneme extractor must emit silence markers. Rhubarb emits X cues for silence (mapped above to viseme_sil); cloud Whisper-phoneme models often don’t — add a silence-detection pass if you go that route.
Lipsync drifts behind audio. Audio playback latency on some platforms is 30–60ms. Either delay the audio start by 1 frame, or pre-shift the animation by the platform’s known latency. On desktop Linux audio output can be 60ms behind; on Steam Deck ~20ms.
Visemes pop on/off. Crossfade between viseme keyframes — the constant is Animation.INTERPOLATION_LINEAR (there is no TRACK_INTERPOLATION_LINEAR; the real set is INTERPOLATION_NEAREST / _LINEAR / _CUBIC / _LINEAR_ANGLE / _CUBIC_ANGLE). Apply it per track with anim.track_set_interpolation_type(idx, Animation.INTERPOLATION_LINEAR) and ensure each viseme’s weight ramps from previous to current. Default bake does this; if you wrote custom keyframes with NEAREST interp, switch.
Smile fights lipsync. Both write to mouth BlendShapes. Solve by additive layering: lipsync layer outputs deltas from neutral, smile layer outputs deltas from neutral, sum them, clamp 0..1. ARKit shapes are sum-safe up to ~1.5; clamp prevents over-rotation.
Eyes look dead. Lipsync is mouth-only; without blinks and saccades the face is uncanny. Add an idle blink track (every 4–8s, jittered) and a small saccade track (random eye movement up to 5°). Both can be one-shots fired by a Timer.
Phoneme extraction returns gibberish. Audio is too noisy, mismatched language code, or compressed too aggressively. Re-export the source as 22kHz mono WAV before sending. Don’t lipsync from a 64kbps MP3.
Custom rig has different BlendShape names. The bake assumes ARKit-52. If the rig uses Preston Blair or a custom set, write a name-mapping Dictionary and apply it during bake. Only do this once per character.

Quality bar

~12 phonemes/sec is normal English speech; the bake produces ~25 keyframes/sec across all tracks. Below 60fps playback on weak hardware: enable BlendShape track interpolation = NEAREST (loses smoothness, gains 30% perf). Reserve for mobile / very crowded crowd scenes.
Lipsync alone is ~70% of “alive”. Add idle blink + idle micro-head-bob (procedural-animation) and you’re at ~95%. The remaining 5% is brow articulation tied to dialogue sentiment, which is bespoke per scene.

Anti-patterns

Driving BlendShape weights from _process with hand-coded curves. The Animation track is sample-accurate and re-uses the AnimationTree’s interpolation; manual code drifts and stutters.
Using a single “talking” loop instead of phoneme-driven lipsync. The 1990s look. Always extract phonemes; the cost is trivial.
Forgetting to gate audio + animation on the same frame. If you play() audio in _ready and fire the OneShot in _process, you’ll see ~16ms drift at line start. Both calls in the same function, same frame.
Using lipsync for non-talking sounds (groans, screams). Lipsync needs phonemes; non-verbal vocals confuse the extractor. For grunts, drive jawOpen from the audio’s RMS envelope instead — much simpler.

Edge cases

Multilingual VO. Rhubarb supports English best; for other languages, pass --recognizer phonetic (language-agnostic, less accurate) or use a multilingual Whisper-phoneme model on Replicate / HF. Cross-language lipsync (extract as English on Spanish audio) gives ~70% accuracy — bad enough that subtitles are needed regardless.
Singing. Phoneme extractor handles sustained vowels well, but consonant timing is loose. For sung dialogue, manually keyframe consonants and let extracted vowels fill in.
Aside / muttering at low volume. Phoneme extractor needs > -40 dB. Boost the source clip before extracting if the line is intentionally quiet, then play it at the original volume in-engine.
Stylized character with no jaw bone (e.g., a cartoon ball). No bone, but BlendShapes can still drive a “morph open” shape. Same pipeline; skip the jaw-bone-track and only animate BlendShapes.

Fallback (no Rhubarb)

If Rhubarb won’t install on the user’s platform, run a Whisper-phoneme model on Replicate (e.g., cjwbw/whisper-phoneme) or aeneas-align for forced alignment. Download the JSON, swap the bake function for one that consumes ARPAbet phonemes (table in Reference card). Same end result — replayable BlendShape Animation resource. For projects that cannot use any cloud or third-party tool, Summer Engine’s AudioStreamGenerator with hand-rolled vowel/consonant detection from RMS and zero-crossings gives only a rough stylized approximation. Do not present it as production speech recognition or photoreal lip sync.

Handoff

For voice generation upstream, voice-line (which wraps assets.generate_audio({ capability: "text_to_speech", ... })).
For dialogue scripts and conversation flow, design-npc.
For the AnimationTree this layer composes into, animation-tree.
For idle blinks, saccades, and head-tracking that complement lipsync, procedural-animation.
Full performance capture (face and body together) is out of scope; use an external capture pipeline.

From a shell

No Summer MCP tools in this session? Run each one as a command; when summer --version is missing or below 4.0, use npx -y summer-engine@latest in place of summer. Fields are flags in kebab-case (--scene-path res://main.tscn) or one --input '<json>', and summer ops describe <op> prints the exact input. The summer-cli skill has the rest.

In this skill	Command
`assets.generate_3d`	`summer assets generate-3d`
`assets.generate_audio`, `summer_assets_generate_audio`	`summer assets generate-audio`
`summer_assets_generate_motion`	`summer assets generate-motion`
`summer_assets_search`	`summer assets search`
`summer_add_node`	`summer engine add-node`
`summer_get_script_errors`	`summer engine get-script-errors`
`summer_inspect_node`	`summer engine inspect-node`
`summer_inspect_resource`	`summer engine inspect-resource`
`summer_run_script`	`summer engine run-script`
`summer_save_scene`	`summer engine save-scene`
`summer_set_prop`	`summer engine set-prop`
`summer_set_resource_property`	`summer engine set-resource-property`
`summer_jobs_wait`	`summer jobs wait`

character-animation-wiring: Wire a rigged character end to end: real clips and bones, locomotion states, method tracks, root motion.
design-npc: Design enemy, NPC, boss or companion behavior: perception, state machine, telegraphs, script stub.
animation-tree: Wire AnimationTree state machines and blend trees: locomotion blends, attack interrupts, hit reactions.
generate-motion: Attach a library animation clip (idle, walk, run, attack) to a rigged character via AnimationPlayer.
procedural-animation: Runtime bone changes on top of clips: head look-at, foot IK, hand grabs a prop, lean, recoil.
retarget: Apply animation clips from one rigged character to another: one clip library, many models.

Source: library/skills/facial-and-lipsync/SKILL.md, rendered from summer-engine@4.3.0 on npm.

​When your agent uses it

​Install

​The skill

​When to use this skill

​When NOT to use this skill

​Steps

​1. Confirm the head has BlendShapes

​2. Get the audio

​3. Extract phonemes (the generative step — runs OUTSIDE Summer)

​4. Persist the viseme track as an AnimationLibrary entry

​5. Wire into the AnimationTree

​6. Add the emotional layer (authored, not generative)

​Confirmation gates

​Reference card

​ARKit-52 viseme subset (the 15 that matter for English lipsync)

​Phoneme → viseme mapping (for hand-rolled extraction)

​Pitfalls

​Quality bar

​Anti-patterns

​Edge cases

​Fallback (no Rhubarb)

​Handoff

​See also

​From a shell

​Related skills