Most voice-AI vendor demos run at sub-700ms response latency, 97% transcription accuracy, and a single clean speaker reading scripted prompts in a quiet room. None of that is what an enterprise call looks like. The buyer that signs a procurement order then deploys into a 55–65 dB contact centre, a tenant calling from a Birmingham bedsit with the kettle on, or a clinical team using terminology no general-purpose model has ever seen.
The gap between demo and production is where pilot programmes die. By April 2026, median end-to-end voice-AI latency in production hit roughly 680ms — down from ~1,200ms in 2024, but still over double the 200–300ms human conversational window where listeners stop noticing the delay. And that is the median: under noise, accents, and barge-in, real systems drift into the 1,400–1,700ms range that empirically breaks deals.
This guide is shipped by the team behind Dilr Voice — enterprise voice AI live across 40+ countries with sub-300ms median latency and UK-English tuning out of the box. Or see DATS, our five-stage AI consulting system for regulated buyers.
This post is the buyer-side benchmark sheet. It is not a vendor leaderboard. It tells you what to measure before you sign, what those numbers should be under enterprise call conditions, and how to write the procurement test that surfaces the gap. The pillar it sits under — the enterprise AI voice agents guide — covers the wider architecture; this piece is the measurement layer underneath it.
The benchmark that matters is not demo-room accuracy. It is performance under the four conditions every enterprise call surface contains: background noise, regional accents, mid-call interruptions, and domain-specific vocabulary. Any vendor whose benchmark documentation omits these four is benchmarking marketing, not infrastructure.
- Demo ASR: 97%+ in clean audio. Production ASR in contact centres: 88–93%.
- Demo latency: sub-700ms. Production median across the market: ~680ms. P95 under barge-in: 1,400–1,700ms.
- Background noise at 55–65 dB cuts transcription accuracy by 15–30% without noise-robust models.
- Real-world ASR spans 91–98% depending on environment — a 7-point gap that decides whether the agent transfers or hallucinates.
The four conditions that decide whether a voice agent is enterprise-ready
Vendor benchmark pages typically lead with two numbers: end-to-end latency in milliseconds and Word Error Rate (WER) as a single percentage. Both are real signals; both are also game-able to the point of being misleading. A 600ms latency figure from a vendor demo tells you the infrastructure ceiling under perfect conditions. A 97% WER figure tells you how the model handles a single American speaker reading the Wall Street Journal. Neither tells you whether your customers will get answered.
The four conditions that matter in a procurement test:
1. Background noise. Most enterprise call surfaces — contact centres, retail back offices, field-service trucks, NHS reception desks — sit at 55–65 dB of ambient noise. That noise floor reduces transcription accuracy by 15–30% without noise-robust models. The gap shows up not as a slightly lower WER but as a different failure mode: the agent mis-extracts a postcode, mis-hears a yes/no on a payment confirmation, or transcribes a competitor brand name as a generic noun. Every one of those is a customer-experience or compliance incident.
The benchmark you ask for: WER and intent-accuracy delivered with a synthetic 60 dB contact-centre noise track mixed into the test audio. Vendors who haven't run this evaluation will say it's "outside our standard benchmark methodology." That is the answer.
2. Regional accents and code-switching. UK enterprise call estates contain Glaswegian, Welsh-English, Multicultural London English, Indian English, Polish-English, and a long tail of code-switching speakers. Most public ASR benchmarks report a single English figure that hides 5–10 points of variation by accent. Independent evaluation of the speech-to-text market shows English ASR in low-noise conditions has improved from ~95.1% (2023) to 97.3% (2026) — but those numbers collapse outside North-American General English.
The benchmark you ask for: WER segmented by accent, with at least five UK-relevant accent categories. If a vendor cannot produce this, your real-world performance will be one or two standard deviations below their marketing claim. We covered the wider framework for this in voice AI accuracy: beyond WER for enterprise.
3. Mid-call interruptions (barge-in). Human conversation operates inside a 200–300ms response window. Exceeding it triggers neurological stress responses that break conversational flow — listeners interrupt, repeat themselves, or hang up. The hardest single test of a voice agent is whether it gracefully handles the customer interrupting the agent — the barge-in case. P50 latency numbers say nothing about this; you need P95 latency under barge-in, measured end-to-end including TTS cutoff and dialogue state recovery. We unpack this gate in voice AI barge-in handling: why interruptions break deals.
4. Domain-specific vocabulary. A general-purpose ASR model trained on web audio does not know your product SKUs, your drug brand names, your claim-reference number formats, your borough codes, or the name of the loyalty tier you renamed last quarter. Domain vocabulary is the single biggest contributor to invisible error: the agent transcribes confidently but wrongly, and downstream systems act on the wrong value. The benchmark you ask for: WER on a custom 200-term lexicon drawn from your own call recordings, before and after the vendor's domain-tuning workflow runs.
The benchmark sheet enterprise buyers should actually demand
The same diagnostic logic underpins our AI placement diagnostic — a fixed-fee assessment used before any deployment commitment, deliberately designed to surface the production-condition gaps that demos hide. Here is the benchmark sheet we run against every vendor before recommending one.
The architecture decision underneath these numbers — pipeline composition, streaming vs. batch, where the LLM call sits — is covered in LLM vs scripted voice agents: enterprise guide and voice AI orchestration vs platform: the enterprise choice. Strategy buyers should also read the cross-cluster piece on the enterprise voice AI vendor checklist and AI voice platform: enterprise selection criteria, which sit alongside this measurement layer in any serious procurement.
Every box adds latency. Parallel pipelines — where TTS begins streaming the first phonemes while the LLM is still generating the tail of the response — are how the fastest production systems hold sub-500ms end-to-end. Sequential pipelines that wait for full LLM completion before starting TTS sit at 1.4–1.7s. The architecture is the benchmark; the millisecond figure is the symptom.
The procurement test table
| Condition | Demo benchmark | Enterprise production target | Failure mode if missed |
|---|---|---|---|
| Latency, P50 quiet | < 700ms | < 700ms | None — table-stakes |
| Latency, P95 with barge-in | Not reported | < 1,100ms | Caller hangs up, agent finishes alone |
| ASR, clean English | 97%+ | 97%+ | None — table-stakes |
| ASR, 60 dB noise mixed | Not reported | > 92% | Wrong intent extracted, downstream action wrong |
| ASR, 5 UK accent categories | Not reported | > 90% per accent | Bias incidents, regional CX collapse |
| Domain vocabulary, 200-term | Not reported | > 95% post-tuning | Silent data-quality contamination |
| Concurrent-call accuracy | Not reported | No degradation at 500 concurrent | Black-Friday-style collapse |
A vendor that cannot produce numbers for the bottom four rows is asking you to underwrite their production performance with your own brand. We run an enterprise voice AI deployment platform tuned for UK-English call conditions; this same sheet is how we benchmark ourselves internally before any client cutover.
What this means for the buyer
Two practical implications.
First — write the test into the contract, not just the RFP. Vendor RFP responses can claim anything. Acceptance criteria written into the master services agreement — with the seven rows above attached as Schedule A and a measurement methodology agreed — are what give finance leverage when production reality diverges from the demo. This is the procurement discipline missing from most pilots that end up in AI voice pilot purgatory: why 70% of programmes stall, and the muscle our AI operating model consulting practice installs inside enterprise voice programmes.
Second — accept that some of the gap is yours, not the vendor's. Custom vocabulary, contact-centre acoustics, and accent distribution are your data. A vendor will tune to them, but only if you supply representative samples. Enterprise buyers who treat domain-tuning as the vendor's problem ship pilots that hit production limits on day three. The vendor evaluation methodology covered in our voice AI agent QA and testing framework is the buyer-side counterpart to this benchmark sheet — together with our AI execution office, they form a complete pre-procurement gate.
Want to see this in production? Try Dilr Voice live (free, $20 credits), book an AI placement diagnostic on your real call audio, see our DATS methodology, or speak to the operators who run these benchmarks every week.
The number that matters in May 2026 is not how fast your voice agent answers in a demo. It is how often it answers correctly when the line is bad, the caller has an accent, the noise floor is 60 dB, and your product names are not in any pre-trained vocabulary. Measure that — and write the answer into procurement.
Benchmark your shortlist on real call conditions.
30-min scoping call · No deck · Confidential. We will tell you which vendor benchmarks survive your real call estate, and which ones collapse on contact.
Written by the Dilr.ai engineering team — practitioners who ship enterprise AI in production. Follow us on LinkedIn for shipping notes, or subscribe via the RSS feed.
Sources: Hamming AI — Voice AI Latency benchmarks 2026 · Speechmatics — Voice AI in 2026: 9 numbers.