Griffin ranks second on LSE-C: how to judge lip sync yourself

What the benchmark footnote says

Tavus's Griffin page, read 2026-10-03, lists Griffin first on DOVER, FID and THEval and second on the lip-sync metric LSE-C, with a score of 7.27. Tavus explains the second place: it says LSE-C rewards pronounced mouth movement beyond what looks natural.

That is a useful caution for any buyer. A metric that scores sync can favour exaggerated mouths, so a high number does not prove that a face looks right. Your own eyes, on your own clip, still decide.

A check that uses timed stills

Sume's video inspect returns a word-timed transcript when transcribe: true and stills at times you name with frames.at, up to 24 per call. The transcript's words[] give each word's time, so you can pull a still at the start of a plosive such as p, b or m and see whether the lips are closed there.

First import the clip with POST /v1/media-imports if it is not already on media.sume.com. Run the transcript once, pick four or five words, then run a second inspect with frames.at set to those times.

curl -X POST https://api.sume.com/v1/video-inspect \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: lipsync-check-001" \
  -d '{
    "video_url": "https://media.sume.com/artifacts/artf_demo/avatar.mp4",
    "frames": {"at": [0.4, 1.1, 2.3, 3.0], "format": "png"},
    "transcribe": true,
    "language_code": "en"
  }'

What to look at

• Lips closed on p, b and m sounds, open on vowels.
• Teeth that stay the same shape from frame to frame.
• The mouth not opening wider than the word needs; an over-wide mouth is what a sync metric can reward.
• Frames near cuts in a multi-scene clip, where a seam can show.

Cost and limits

The transcript is billed at a public $0.01 per audio minute on top of the inspect's compute; the docs tell you to confirm it live in the catalog. A silent clip returns inspect_source_has_no_audio, so check probe.has_audio first. An at value past the clip's length returns frame_time_out_of_range.

A small routine

Run the check on the first clip of a series, not every clip. If the avatar passes on plosives and the mouth shape looks steady, the same avatar and quality tier will behave similarly on later clips. Re-check when you change tier, language or aspect ratio.

Write down what you saw next to the job id. A note such as "closed on b at 1.1 s, teeth steady" is more useful in six months than a feeling.