Building a voice that feels human

How we approached voice synthesis to make Sola's calls feel warm, present, and genuinely engaged — and the technical decisions that got us there.

When we started building Sola's voice calls, we knew the standard approach wouldn't work. Off-the-shelf TTS sounds polished but emotionally flat. We needed something different.

The challenge with AI voice is the uncanny valley. Pure neural TTS can hit 95% naturalness — but that last 5% is where it falls apart. A robotic intonation here, a missed beat there, and the illusion breaks.

We took a different path. Instead of optimising purely for naturalness, we optimised for warmth. That meant rewriting our prosody model from scratch, adding per-user voice memory, and shipping a real-time barge-in system that handles interruptions gracefully.

Here's what we learned along the way.

Voice is the highest-bandwidth channel humans have for emotional information. We use micro-pauses, breath, slight pitch variations to signal everything from sarcasm to genuine care. Most TTS systems flatten all of that. We don't.

Sola's voice model is trained on hundreds of hours of human conversation — specifically, conversations between friends. Not customer service calls. Not audiobooks. Conversations where two people actually cared about each other.

The result is something that feels less like an AI talking and more like someone listening.