01.1 — Feature· 6 min read

AI video analysis.

Before Klyph cuts anything, it watches and listens. Analysis is the layer that finds what matters — the hook, the payoff, the quiet beat that still works on mute.

§ 1

What gets analyzed

Klyph runs four passes over every source:

  • Transcript pass — diarized speech-to-text with speaker labels, word-level timestamps, and filler detection.
  • Semantic pass — topic segmentation and narrative arc detection across the whole transcript.
  • Visual pass — scene detection, face tracking, and motion intensity sampled every 250ms.
  • Audio pass — energy, laughter, pauses, and emphasis per utterance.
§ 2

How a moment gets scored

Each candidate segment is scored against a blended model that weights retention curves from millions of real short-form posts. The output is a score between 0 and 100, plus sub-scores you can inspect.

FieldTypeDescription
hook0–100Strength of the first 3 seconds
coherence0–100Does the clip stand alone without context?
emotion0–100Expressiveness of voice and face
pacing0–100Density of information over time
payoff0–100Is there a resolution or punchline?
Tip
Override the weights for your niche.
On paid plans you can create a scoring profile that biases any sub-score. A meditation channel, for example, inverts pacing so calm moments rank higher.
§ 3

Latency & limits

Analysis runs on dedicated GPU pools — expect ~0.6× real-time for 1080p sources. A 60-minute podcast typically returns ranked candidates in under four minutes.

§ 4

Languages

Transcription supports 42 languages. Semantic scoring is production-quality in English, Spanish, Portuguese, French, German, Hindi, Japanese, and Arabic, with the remaining languages in preview.