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.
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.
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.
| Field | Type | Description |
|---|---|---|
| hook | 0–100 | Strength of the first 3 seconds |
| coherence | 0–100 | Does the clip stand alone without context? |
| emotion | 0–100 | Expressiveness of voice and face |
| pacing | 0–100 | Density of information over time |
| payoff | 0–100 | Is there a resolution or punchline? |
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.
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.