Resources

Read before you buy.

We publish the material we would want as a buyer: benchmarks with the test set attached, an ROI model with the assumptions exposed, migration checklists with the parts that hurt highlighted, and post-mortems from calls that went wrong in production.

Book a 30-minute demo See the gap analysis vs LivePerson
▪ 19 documents in the library ▪ 3 post-mortems, unedited ▪ No email wall on any of it
Start here
Benchmark: latency and accuracy, six vendors

Same 1,000-call set, same carrier, human-judged. Updated 9 June 2026.

 Post-mortem: the postcode we misheard

34 delivery calls, one digit, five weeks of guardrail work.

 Glossary: 14 terms without the fog

Barge-in, WER, warm transfer, data residency — defined in one sentence each.
Featured

Three documents that answer real procurement questions.

Everything in this library is written by the team that builds the runtime, not by a content agency. Where a number is bad for us, it stays in.

REPORT · 42 PAGES · 9 JUNE 2026

Voice Agent Benchmark — Q3 2026

Six vendors, one test set, one carrier. We measured turn latency at p50 and p95, task completion, entity accuracy on noisy lines, fabrication rate and word error rate on accented English.

Nivākya ranked first on latency and task completion, third on fabrication rate behind Kestrel CX. The methodology section lists every call.

Open the tables
GUIDE · 68 PAGES · 24 JULY 2026

The Rehearsal Playbook

How to declare behaviours as testable rules, run a crowd of synthetic callers against a draft agent, read the failure list, and produce a signed version your risk committee will accept.

Includes the 41 rules we start every collections deployment from, and the three we deliberately leave out until the second release.

See Rehearsal
MODEL · 34 INPUTS · REV 11

Cost-per-contact model

A working model that compares a BPO contract, a voice-agent deployment and the hybrid most of our customers actually end up running. Every assumption sits on a visible tab.

Defaults are drawn from 61 production workspaces and are deliberately pessimistic: 25% automation ceiling, 1.9 attempts per contact.

Compare plans
The library

Nineteen documents, four shelves.

Benchmarks

Six vendors, one test set, one carrier.

Every vendor was given the same 1,000 recorded calls across three regions and asked to complete the same tasks: book a visit, confirm an address, take a payment promise, escalate to a human.

Vendor Median turn latency p95 latency Task completion Entity accuracy, noisy line Fabrication rate WER, accented English
Nivākya v4 0.61s1.38s92.4%96.1%0.4%6.2%
Kestrel CX Voice 0.98s2.44s85.3%91.2%0.3%9.4%
Aspen Dialog 1.12s2.90s83.1%89.4%1.9%10.6%
LivePerson Conversational AI 1.24s3.10s81.0%88.7%2.6%11.9%
Northvoice Cloud 1.41s3.86s78.6%85.9%3.1%13.2%
Ensemble Voice 1.77s4.20s74.2%83.5%4.2%15.1%
Methodology, in full. Calls were replayed from a consented corpus of 1,000 English and Hindi-English conversations recorded between February and April 2026. All six vendors were reached over the same carrier on the same day-parts, in uk-south, eu-west and ap-south. Latency is measured at the carrier edge as the interval between the end of the customer's utterance and the first audio byte of the reply. Fabrication means a claim made by the agent that no tool result, knowledge-base article or customer utterance supports.
What we did not control. Task completion is judged by two human raters with a 0.91 agreement rate; disagreements were resolved by a third rater who did not know which vendor produced the call. Aspen and Ensemble ran on their default US voices for the Hindi-English segment, which is not their strongest configuration and we say so in the appendix rather than quietly removing them.

Raw transcripts, per-turn timings and the rater sheets are available to customers and to any prospect who has signed an NDA. Request them through the demo form.

Post-mortems

Three calls we got wrong, written up in public.

We publish these because a vendor with no failure record is either very young or not looking. Each one lists what broke, how long it lasted, and the specific change that followed.

INCIDENT 2026-05-11 · 34 CALLS

The postcode we misheard

What broke. On a noisy line, ASR returned “ZR-4” for “ZR-14” on a delivery-rescheduling agent. The agent read the address back as a single string, the caller said “yes” to the wrong thing, and 34 reschedules went out with an undeliverable address.

What changed. All alphanumeric fields now require digit-by-digit readback with explicit confirmation per character, and a keypad fallback fires after the second failed readback. Median handling time rose 4.1 seconds. We accepted that.

INCIDENT 2026-02-02 · 3 CALLS

The retry ladder that kept dialling

What broke. A caller asked us to stop calling, in the words “please don't ring this number again”. The opt-out classifier only matched a fixed set of phrases and missed it. The ladder dialled twice more over four days.

What changed. Twenty-two opt-out phrasings added, suppression is now immediate and irreversible for 90 days, and every suppression event goes to a human queue for next-day review. The classifier's job changed from “is this an opt-out” to “could this plausibly be an opt-out”.

INCIDENT 2026-03-27 · 18 MIN, 1,240 CALLS

The policy engine stall in ap-south

What broke. A rule-set publish evicted a warm cache faster than it could refill, and policy evaluation stalled for 18 minutes in one region. Calls degraded to a static script that completed bookings but lost every upsell branch.

What changed. Publishes are staged to 5% of traffic for ten minutes, a p95 regression on that slice triggers automatic rollback, and cache prewarm is now a required step in the release checklist rather than an optimisation.

How we write these. Within 72 hours of a customer-impacting incident we publish a summary to affected workspace owners, with the call count, the affected region and the change list. The three you just read are the ones where the fix was expensive enough that we wanted other operators to learn from it. Status history lives on the status page.
Newsroom

Product and company news.

12 AUG 2026Nivākya opens a Dubai office and hires a regional lead

Gulf deployments now run from in-region infrastructure with Arabic dialect models for Gulf, Levantine and Egyptian speech.

 24 JUL 2026Rehearsal 3.0 ships 120 adversarial personas

The persona library now covers background noise, third-party callers and deliberate prompt-injection attempts.

 09 JUN 2026Voice Agent Benchmark Q3 2026 published

Six vendors, one test set. Nivākya leads on latency and completion, third on fabrication rate.

 21 APR 2026Series B: £48M led by Ridgeline Capital

Funding is earmarked for regional infrastructure and a second India engineering centre in Pune.

 03 MAR 2026SOC 2 Type II renewed with zero exceptions

The audit window covered 12 months and all four regions, including the Bengaluru support desk.

 15 JAN 2026Hindi-Urdu code-switch model enters general availability

Word error rate on intra-sentence switches fell from 14.8% to 7.1% on our internal eval set.
Events

Four sessions before the end of the year.

All of them are recorded, none of them are gated, and each one ends with thirty minutes of questions that we do not pre-screen. Bring the awkward ones.

Reserve a seat Listen to the podcast instead →
08 OCT 2026 · 10:00 ET · WEBINARRehearsal before rollout: a collections agent in 14 days

With the Head of Collections at a mid-market NBFC who shipped in thirteen. Includes the two rules her risk team rejected.

22 OCT 2026 · 16:00 GMT · TEARDOWNLine-by-line: an after-hours booking agent

We open a production rule set on screen and explain every guardrail, including the ones that cost us conversions.

06 NOV 2026 · 11:30 IST · PANELMultilingual voice in India: accents, consent, cost

Three operators on one panel, disagreeing publicly about whether one number should carry five languages.

19 NOV 2026 · 09:00 PT · OFFICE HOURSBring your rule set, we'll try to break it live

Sixty minutes, six volunteers, no slides. We run adversarial personas against whatever you paste in.

Podcast — Signal Path

Conversations with the people who rebuilt their voice channel.

Forty minutes, one operator, no vendor pitch. Transcripts are published within a week, and guests get to correct their own words before we ship them.

EpTitleGuestLengthPublished
14Nobody misses a queue if nobody notices oneVP Customer Experience, national property developer38 min02 Sep 2026
13The ninety seconds that decide a collections callHead of Collections, regulated NBFC44 min12 Aug 2026
12Buying voice AI without a sandboxProcurement Lead, six-hospital network31 min22 Jul 2026
11Transcripts as evidenceCompliance Director, general insurer46 min01 Jul 2026
10Your IVR tree is the best documentation you haveDirector of Service Operations, telecom29 min10 Jun 2026
09Hindi, Tamil, English: one numberHead of CX, insurance distributor41 min20 May 2026
08The rehearsal report that killed a projectProgramme Manager, retail bank36 min29 Apr 2026
07Agent handover without a sighTeam Lead, outsourced contact centre33 min08 Apr 2026

Signal Path is also available as a written digest — every episode ships with a two-page summary and the three timestamps worth hearing. Ask for the digest through the demo form.

Elsewhere

Community, partners and the people who grade us.

COMMUNITY

The Signal Room

A peer forum for operators running voice agents in production. 4,100 members, roughly 62 answered threads a week, one live clinic per month.

  • ▪ Rule sets shared under an open licence
  • ▪ Monthly “break my agent” clinic, recorded
  • ▪ Direct line to Nivākya engineers on Thursdays
Request an invite
PARTNERS

46 implementation partners

Systems integrators and CX consultancies that build on Nivākya, audited annually on delivery quality rather than on revenue.

  • ▪ 12 Premier partners with certified architects
  • ▪ 27 Select partners, regional coverage
  • ▪ 7 technology partners shipping on the API
Partner with us
ANALYSTS

Third-party evaluation

We submit to independent review twice a year, including the categories we tend to lose. Reprints are available at no cost.

  • ▪ Realtime Voice Quadrant 2026 — Leader
  • ▪ Voice Automation TCO Study — second on cost at scale
  • ▪ Conversational AI in EMEA — Challenger on localisation
Read our gap analysis
Glossary

Fourteen terms, one sentence each.

Voice AI has borrowed its vocabulary from three industries and defined it in none of them. Here is the version we use internally.

ASR
Automatic speech recognition — the stage that turns audio into text, and the stage most likely to fail on an accented or noisy call.
TTS
Text-to-speech — synthesis of the agent's voice, judged on naturalness, prosody and how quickly it can start after a decision.
Barge-in
The customer talking over the agent. Good agents stop within about 150 milliseconds without clipping the customer's first word.
Turn latency
The interval between the end of the customer's utterance and the first audible byte of the agent's reply. Measured at the carrier edge, not inside our own network.
WER
Word error rate — substitutions, insertions and deletions as a share of the words actually spoken. Below 8% is workable on accented English without a prompt.
VAD
Voice activity detection — the decision about whether a sound is speech, noise or silence. Get this wrong and the agent either interrupts or waits politely forever.
Tool call
An action the agent takes outside the conversation: checking a calendar, writing a CRM record, sending a payment link. Every tool call is logged with its arguments.
Warm transfer
Handing the call to a human with context attached — transcript, caller history and the reason for escalation — instead of dropping the caller into a fresh queue.
Concurrent call
A single call in progress at one moment. Your carrier trunk limit, not our platform, is usually the binding constraint.
Rehearsal run
A batch of simulated calls against a draft agent version — up to 10,000 — producing a pass or fail against behaviours you declared in advance.
DNC list
Do-not-call register. Nivākya checks regional registers plus your own suppression list before a number is ever dialled.
PII redaction
Removing names, card digits, addresses and identifiers from a transcript before it is persisted. Redaction happens in the pipeline, not in a report afterwards.
Data residency
The physical region where audio, transcripts and derived analytics are stored. Pinned per workspace and not changed without a signed instruction.
Intent score
A 0–100 estimate of how close a caller is to taking the action you care about, with the sentences that produced it attached.
Start

Bring your hardest question to a 30-minute call.

Most of these documents exist because a buyer asked us something we could not answer at the time. Keep asking — it is how the library gets better.

Book a 30-minute demo Read the LivePerson gap analysis

No email wall, no gated PDF, no report that turns out to be a brochure.