Verifying Identity with Voice and Face Recognition: Building Sparsha's Onboarding

The trust problem behind every dating and streaming app

Sparsha combines dating, calling, live streaming, room streaming, and reels posting in one app. Every one of those features depends on the same thing: users trusting who they are actually talking to. A platform that lets anyone claim any identity at sign-up has a safety problem baked into its foundation, no matter how good the rest of the product is.

Sparsha is live on Google Play, used by hundreds of real users, primarily in India. The client's requirement was direct: verify a user's gender before they could interact with anyone else on the platform, using more than a self-reported checkbox.

Why one verification method is not enough

A photo alone can be borrowed. A voice sample alone can be spoofed. We built the onboarding flow to require both voice analysis and facial recognition, so a user has to pass two independent checks rather than one that is comparatively easy to fake.

This runs as a gating step in onboarding itself. A new user cannot reach the dating, calling, or streaming features of the app until verification clears, which keeps the safety check ahead of the interaction rather than bolted on after something goes wrong.

What we actually built

Sparsha is a single Flutter app for Android and iOS backed by a Node.js API, all behind the same verified-identity gate. Building this as one unified app, rather than separate apps per feature, mattered for the verification model specifically: identity only needs to be confirmed once, and every feature downstream can trust it.

• One-on-one and group calling
• Live streaming and multi-host room streaming
• Reels posting and discovery
• The dating and matching flow itself

Lessons for anyone building identity verification into an app

A few things held up across this build that are worth knowing if you are scoping something similar.

• Verification should block access to sensitive features, not just log a warning
• Combining two independent signals, voice and face, is meaningfully harder to spoof than either alone
• Verification has to run fast enough that it does not feel like a punishment for signing up, or users abandon before they finish