Fine-tuning
Ten recordings. Two hours. Production-ready.
Upload annotated call recordings and OneVox adapts vocabulary, pacing and emotional register to your vertical — automatically. No prompt engineering. No manual tweaking.
How it works
Drop in call recordings (WAV, MP3 or a Twilio / Genesys export) with optional outcome labels like resolved, escalated or CSAT.
Our pipeline learns your vocabulary, pacing, pause length and emotional register — per vertical, per intent.
Every fine-tune is scored against a held-out split and your current production model before it can ship.
Route a slice of live traffic to the new voice profile and compare resolution, handle time and CSAT side by side.
Promote the winner to 100% in one click. Each version is immutable and tagged, like a release.
Something off? Roll back to any previous version instantly — no redeploy, no dropped calls.
One command, or one click
$ onevox finetune create \--agent aria \--recordings ./recordings/2024-q4 \--vertical support \--eval-split 0.2✓ 10 recordings ingested · 6h 12m of audio✓ Prosody profile trained · support-empathy-v3✓ Eval passed · CSAT +0.6 vs baseline→ Deployed aria v2.4.1 to 10% of traffic
What gets tuned
Product names, SKUs, medical or legal terms and local place names are pronounced the way your customers say them.
Speaking rate and pause-after-question are learned from your best calls, not guessed.
Empathy, urgency and warmth adapt to intent — calm for billing disputes, upbeat for sales.
Barge-in sensitivity and back-channel cues (“mm-hm”, “got it”) match your callers' rhythm.
Safe by default
Recordings are encrypted at rest and never used to train shared models.
Names, card numbers and addresses are redacted before training begins.
Every model is versioned, diffable and rollback-safe, with a full audit log.
Send us 10 recordings and we'll show you the tuned agent in a 30-minute session.