Hundreds of teams are building voice agents. Very few build what those agents run on. Voicing is the runtime underneath: speech in, decision, action, speech out, inside your firewall.
Past simple questions, every extra step adds tokens, tools and latency. The failure is architectural, not a model choice, and it shows in four places.
In the demoIn production
End-to-end latency
About 700 ms1,400 to 1,700 ms
Entity capturePhone, card, date of birth
Feels accurate18 to 34% error
Function calling
90%+ on textUnder 70% on 8+ tools
Cost per minute
$0.08 to $0.1440 to 50% higher
An application can be swapped in a quarter. The layer underneath takes a security review, a telephony integration and a model validation to replace. Choose the runtime first.
Four products ship on it. Or build your own on the SDK.
Every layer is researched, trained and served by Voicing. Wrappers fail exactly where enterprise value lives: in the hops between rented parts.
Rented, part by part
A hop between every part, and latency in every hop.
Illustrative sequence
The caller stops speaking.
Endpoint detected60 to 120 ms
Final transcript150 ms
Routed to the action model
Read through, nothing copied out
First audio38 ms
A silence timer
A rented stack, part by part
System of record
700 ms
80th percentile, banking, on-premises, real telephony
1,500 ms
The rented stack answers
Three times under the industry average.
Illustrative turn. Marked figures are measured.
0 ms400 ms800 ms1,200 ms1,600 msHumanNaturalNoticeableBrokenHuman to 400 ms, natural to 800 ms, noticeable to 1,200 ms, broken beyondA small model trained through the voice gap, on production audio and graded tool calls, beats larger general models at the moment of value.
It completes the call
97%Voicing action model
One hundred calls each. Hollow is a failed call.
Voicing action model
97%Claude 4.5 Sonnet
91%GPT-4o
84%Llama 3.3 70B
82%One in elevencalls fails at the moment of value
It hears it right
5.8%Open ASR average
Word error rate. Lower is better.
LibriSpeech clean
LibriSpeech other
Open ASR average
Three recognisers, one answer, lower error than any of them alone
It serves it cheaply
$0.013a minuteagainst $0.14for a stack of APIs
Each block is one second of audio, decoded.
0.02Xon GPU
Fifty seconds of audio, decoded every second
0.05Xon CPU
Quantised, on-premises, with no GPU required
Accuracy that only makes commercial sense when you own the models
Every turn is recorded, stage by stage. Every regression is caught overnight, before a caller meets it.
Entity curves
Per slot, language and line
32slotsStage latency
Traced at p50, p95 and p99
15dimensionsSchema audits
Every tool call is graded
100%of tool callsData flywheel
Corrections feed the retrain
NightlyCall quality
Loss, echo and silence, live
RealtimeSemantic outcomes
Completion, escalation, retry
Per journeyIn banking and healthcare the security review is the sales cycle. Policy is enforced in the runtime: one code path for consent, redaction, residency and audit.
A runtime contract grows inside the account. Expansion was the customer’s decision in each of these, taken on the figures from the first weeks.
A tenth of inbound volume to a quarter, in six weeks. At go-live, After expansion, Intents live.
Equifax
2.5Xdaily calls handled
No re-integration, no change window
An IVR replacement that now runs the outbound book. Inbound, Outbound. 6.9K, 32.9K, 48.2K, 44.2K, 78.5K, 125.9K Feb, Mar, Apr, May, Jun, Jul
UTI Mutual Fund
441Kminutes served
Outbound on the same runtime, no new integration
Interpreter queues to real-time multilingual service. Pilot, Expansion.
Great American Insurance Group
4 to 14languages
Interpreter queue removed from the call path
A Spanish-first pilot to ten times the capacity. Pilot, Scaling, Production.
Telmex
10Xcapacity growth
English and Spanish live, same runtime
A working session with an architect: your telephony, your systems of record, your security review. You leave with a deployment outline.
Seven owned layers, running inside your perimeter
