Hebrew Voice Report 2026

TL;DR

The Hebrew Voice Report 2026 is the first annual reference for how Hebrew-native voice agents perform in production. It is written from the inside of an operating system, not from a survey desk. The numbers come from a year of live call traffic on Link AI’s voice infrastructure across multiple Israeli verticals, paired with primary interviews with the operators who deploy them. The report will publish in late summer 2026 with an open dataset attached. This page exists so you can read what is in it, decide if it is useful, and ask to be told when it ships.

Why this report

Voice AI benchmarks today are written almost entirely about English. Hebrew is treated as a long-tail language: a checkbox in a model card, occasionally tested with a single accent and a handful of broadcast clips. That is not the language we work in. Real Hebrew on a real call is code-switched, full of vendor jargon, spoken at conversational speed by people who do not slow down for a transcriber, and routinely interrupted by background noise that is specifically Israeli — cafe espresso machines, warehouse forklifts, kids in the back seat in Be’er Sheva. A benchmark that does not reflect that environment is not a benchmark of Hebrew voice AI. It is a benchmark of clean Hebrew audio.

We built the deployment that this report measures because we needed it to exist. The report exists because the data we collected to operate that deployment has answers that nothing published anywhere else does. Pickup latency at the 50th, 90th and 99th percentile across the day. ASR accuracy on Mizrahi, Ashkenazi, Russian-influenced, and Arabic-influenced Hebrew pronunciations. Call-completion rates by intent and by vertical. Cost per resolved call at the volumes where unit economics start to matter. We are publishing them because the absence of public numbers is currently the single biggest reason Israeli operators mis-buy voice infrastructure.

What is in the report

1. Setting. The state of voice AI in Israel at the start of 2026, the buyers who are deploying it, the vendors who are selling it, and the gap between what is demoed and what survives production traffic.
2. Latency. Pickup-to-first-word latency measured end-to-end on real PSTN calls. Distribution by carrier, hour of day, and call origin. Why “sub-second latency” in a demo and on a live Hebrew call are different numbers.
3. Hebrew ASR accuracy. Word error rate broken down by speaker dialect, by speaker age, by code-switch density (Hebrew interleaved with English brand names, with Arabic loanwords, with technical jargon), and by background noise environment. Annotated against the public Hebrew ASR benchmarks where they exist, so the gap between lab and field is visible.
4. Conversation completion. What share of calls the agent resolves end-to-end without a human handoff, segmented by call intent (booking, qualification, support, FAQ, payment) and by vertical (clinics, professional services, e-commerce returns, real-estate intake, automotive service). Where the handoff triggers actually fire, in the wild.
5. Operator economics. Cost per call and cost per resolved call at the volumes Israeli SMBs actually run (low thousands of monthly calls to low hundreds of thousands). The unit-economics break-even versus a human receptionist, including the line items most vendors omit.
6. Failure modes. The categories of failure that show up only at production scale. Hallucinated booking times. Carrier-side audio clipping. The specific ways a model degrades when the caller is angry. What we shipped to mitigate each one, and what is still open.
7. Outlook. What we expect to change between now and the 2027 edition. The two model releases we are watching. The regulatory items on the horizon for Israeli voice-AI deployments.

Methodology in brief

The bulk of the data comes from Link AI’s own production voice infrastructure over a rolling twelve-month window ending at the report’s cut-off date. Aggregations are computed across all customer deployments and are anonymised before they are written. No customer is named without explicit, written permission, and no aggregation is published at a granularity that could re-identify a single customer or caller. A second layer of data comes from semi-structured interviews with Israeli operators who deploy voice AI on infrastructure other than ours, conducted under the same anonymity terms. Where a number is sourced externally, the source is cited inline.

The full methodology section, including sampling, exclusion criteria, statistical caveats, and where we believe our results may not generalise, will be published as the front matter of the report and as a separate machine-readable schema.org Dataset distribution.

Publication date

The report is targeted for September 2026. We will not push the date to keep a calendar. If the underlying numbers move materially between now and then we will say so, republish, and date the revision. This preview will be revised in place as the report scope settles.

Author

Ori Ben Simhon is co-founder of Link AI and leads the voice-infrastructure work the report draws on. He writes here in his own voice. Editorial responsibility for every number in the report sits with him.