Skip to content
NewTalk to Basher from your browser — free

Platform · Voice engine

The pause before the answer is the product.

Basher answers at the speed of an attentive human — in Arabic, in dialect — until callers forget it's a machine. Below is the budget that makes that true, and the gate that keeps it true.

Illustration — a recording studio at night, a microphone in its lit glass booth, a gold line becoming a waveform at the mic

The budget

The reply lands before the silence gets awkward.

And we've made it impossible for ourselves to break that promise: any change that slows the reply past the 400-millisecond budget is refused automatically before it ever ships. The check runs on the same instrumentation the live pipeline records — not a parallel stopwatch — so the number you see is the number callers feel.

0100200300400500
300–400ms from your last word to its firstEnforced in CI: a p95 above 400ms fails the build.
363ms

p50 mouth-to-ear — engineering bench, synthetic mode

384ms

p95 mouth-to-ear — same bench, 30 turns, against the 400ms budget

What the loop is made of

  • BARGE-IN · 4 LAYERS

    Interruptions that understand Arabic

    A layered controller commits a barge-in only when voice-activity detection, a sustained-speech gate — 250ms for Arabic — and a backchannel filter all pass. So «نعم», «تمام» and «أها» never cut the agent off, and a genuine interruption gets through fast.

  • GULF · LEVANTINE · MSA

    Three dialect families, first-class

    The pipeline is Arabic-first across Gulf, Levantine and Modern Standard Arabic — down to details like a backchannel set built from how Arabic callers actually speak — with English handled in the same call.

  • ARABIC-NUMERALS

    Numbers spoken as words

    On the self-hosted voice route, Arabic-language agents get every digit run — Latin or Arabic-Indic — spelled out as Arabic words before synthesis, up to 999,999 with gender agreement, so an order number is not read back as English digits. Placeholders and URLs are left untouched; English agents are not rewritten. Hosted deployments do not get this step yet.

  • PHRASE-CACHE

    The instant first word

    Greetings, fillers, confirmations and the consent announcement are pre-synthesized when the pipeline starts and played from cache — skipping the model and the synthesizer entirely on the phrases callers hear most.

  • LATENCY-TRACKER

    Measured turn by turn

    Every turn records transcription, first-token and first-audio-byte marks; per-call arrays persist so percentiles are computed over raw turns, and a p95 over the budget raises a breach signal — the same numbers the analytics screen shows.

Spec sheet — as enforced in code

  • ≤ 400ms p95

    Mouth-to-ear budget

  • p50 363ms · p95 384ms

    Engineering bench — synthetic, 30 turns

  • CI: exit ≠ 0

    On budget breach

  • 250ms

    Barge-in sustain gate, Arabic

  • Gulf · Levantine · MSA

    Dialect families

Bench figures are synthetic-mode engineering measurements that validate the production measurement path — they are not provider or customer-call numbers.

The platform

The rest of the platform

Hear the pause before the answer — with your own ear.

The latency numbers above stay ink until you hear them. Talk to the live agent and notice the distance between your last word and its first.

Talk to Basher liveBook a walkthrough