Audio Signature System
Co-designed a proprietary 10-parameter system to resolve voice inconsistency across multi-segment AI-generated videos, ensuring production-grade audio continuity.
The Challenge
JioCreate generates video with Veo in 8-second scenes, and each scene's narration is synthesised on its own. Stitched together, the narrator's voice drifts: a little higher here, faster there, a different tone in the next scene. Viewers notice, and it breaks the video.
The Solution
- Defined a voice as 10 measurable parameters: pitch, timbre, pace, tone, accent, age, gender, energy, range and clarity.
- Gave each parameter a tolerance. Pitch may move by ±5 Hz for emphasis and pace by ±10 words per minute; timbre must not change.
- Every scene is checked against the signature before it moves on, as a quality gate in the JioCreate pipeline.
The Spec
10
Voice parameters per scene
±5 Hz
Pitch tolerance
±10 wpm
Pace tolerance
8 s
Scene length it checks
Key Decisions
- Parameter-based over sample-based: Defined voice identity through measurable parameters rather than audio samples, making the system reproducible and tunable.
- Quality gate integration: Made audio consistency a hard gate in the production pipeline — videos don't progress without passing the signature check.
Tech Stack
Audio Processing
Python
Google ADK
Gemini