AI Narration & Voice Synthesis

Narration written from what the viewer sees — not what was typed

After PII redaction, the narration stage receives cleaned keyframes and the scrubbed action log. An LLM reads the visual context and writes a voiceover script calibrated to match the video’s duration at a natural speaking pace (~150 words per minute).

The script is written from the viewer’s perspective — explaining what’s happening on screen, highlighting the value of each action, and guiding the viewer through the feature without unnecessary jargon.

What the narration model sees

The narration LLM only sees:

• Cleaned, post-edited keyframes (PII already blurred)
• The scrubbed action log (tool calls and DOM snapshots, with sensitive fields stripped at the log layer)
• The original feature intent and success criteria

It never sees raw recordings, pre-blur frames, or any credentials.

Voice synthesis with ElevenLabs

The generated script is passed to ElevenLabs for synthesis. Voice presets are configurable per job:

• professional-female-us — clear, authoritative, ideal for B2B SaaS
• professional-male-us — confident, measured
• friendly-female-us — warmer tone, good for consumer products
• default — system default

Per-client voice presets can be stored in your client configuration and applied automatically to every job for that client.

Raw script preserved

The .txt voiceover script is included in the final bundle alongside the audio file. If you want to re-record with a human voice actor or a custom TTS voice, the script is ready — no need to transcribe or rewrite from the video.

Why narration comes after post-edit

The ordering is a deliberate security design. Running post-edit first means the narration LLM physically cannot describe blurred content. There’s no retroactive audio redaction needed — because the sensitive content was never narrated in the first place.