Making real-time voice feel natural
Aria Thorne
Head of Voice
A natural voice is the difference between talking at an avatar and talking with one. Here’s how The Garden benchmarks and tunes voice for warmth and low latency.
[Journal]
1. The training foundation
The naturalness of our voice comes from the range of our data. Our voice engine was trained on a rich, consented dataset comprising:
500,000+ consented voice samples.
12 million conversational turns analyzed.
98.4% naturalness rating across diverse voices and accents.
2. Technical benchmarks
To keep voice feeling alive, we use multi-signal fusion. This lets the system blend tone, timing, and emotion instead of reading flatly.
“Natural voice isn’t a feature; it’s the baseline for presence. By tuning for warmth and timing, we cut robotic-sounding responses by 24.3% last quarter.”
— Aria Thorne
3. Key findings
Low-latency response: Live voice now returns in <450ms, so conversation feels immediate, not delayed.
Expressive range: We model subtle shifts in tone so a companion can sound curious, calm, or playful as the moment calls for it.
On-device processing: Sensitive audio is handled locally to ensure privacy and zero data leakage.
THE GARDEN
LIVE AVATARS
REAL-TIME PRESENCE
AVATAR STUDIO
MEDIA GALLERIES
MULTI-AVATAR ROOMS
VOICE & CHAT
PRIVATE BY DEFAULT
Tell us how you want to spend your time, and we’ll open a member space shaped around your avatars, your rooms, and your pace.