Talk to our live AI agent right now —Call our live AI —IN +91 79711 70843
NutaanNutaanVOICE AI
Book a demo
PRODUCT UPDATES

What we shipped, and why it exists.

Almost all of it is live in the product today. Where something is still landing, the entry says so plainly — no version numbers, just the capability, the date, and the problem it solves.

PARTNERSHIP

Sarvam’s Indian-language models are coming to Nutaan

We have partnered with Sarvam AI, who build speech and language models for Indian languages first rather than as an adaptation of English ones. It closes the gap our own callers hit most often: an agent that handles clean Hindi well, Marathi passably, and an accent from two districts over not at all.

  • Recognition trained on Indian speech instead of adapted to it. The difference shows up in names, numbers and places — which is precisely what a booking or a captured lead depends on getting right.
  • Code-switching treated as one sentence rather than two languages. “थोड़ा बताइए ना, कितने का padega” is how people actually talk, and it should not be the thing that breaks the call.
  • Voices that carry an Indian accent natively, rather than an English voice pronouncing Hindi words.
  • Where it stands: the agreement is signed and the integration is underway — this entry is the announcement, not the release note. Setu and every existing engine keep running unchanged, so nothing you have already built needs to move.
  • The same models that make an agent work for a caller in Kolkata make it work for an Indian business selling into Dubai or New Jersey. Bharat first, not Bharat only.
PARTNERSHIPsarvamnutaan ai
LANGUAGES

Setu — a bridge across Asia’s languages

Every speech-to-speech engine we could buy refuses most of the languages Asia actually speaks. Ask one for Marathi or Telugu and it answers with an error rather than an accent; the rest are English-first. Setu exists for that gap — 32 Asian languages spoken natively, 70 in all, from Hindi to Japanese and Bengali to Georgian.

  • South Asia — Hindi, Bengali, Marathi, Tamil, Telugu, Kannada, Malayalam, Gujarati, Punjabi, Urdu, Odia, Assamese, Nepali.
  • East and Southeast Asia — Chinese, Japanese, Korean, Indonesian, Malay, Filipino, Vietnamese, Thai, Khmer, Lao, Mongolian.
  • Central and West Asia — Arabic, Turkish, Hebrew, Kazakh, Uzbek, Azerbaijani, Armenian, Georgian.
  • It follows the caller. On one live call the caller moved between Bengali, Hindi, Gujarati, Odia, Marathi and English, and the agent changed with them each time without being told to.
  • Measured on real calls to a real phone: about 900ms to the first word of the greeting, 395ms to each reply — at roughly a sixth of the per-minute cost of the engines it replaces.
  • Stated rather than hidden: the engine understands these languages far better than it writes them down. For the eleven where we have a measured speech-to-text model, the caller is transcribed with that instead, so the saved transcript is accurate as well as the conversation.
32 ASIAN LANGUAGES · ONE CALL, SIX OF THEMবাংলাहिंदीગુજરાતીଓଡ଼ିଆमराठीENThe caller switched. The agent switched with them.32 Asian · 70 total · ~900ms to hello · ~395ms per reply
DIAGNOSTICS

See exactly how long every caller waited

Two silences decide whether a call feels human, and they break for different reasons. The wait before the agent says hello is the engine starting or the phone line; the wait after the caller stops talking is the model or the voice. Both are now measured on every call, separately.

  • Recorded automatically — nothing to switch on and nothing to send us.
  • The usual wait and the slow one side by side, because a one-second median with a three-second tail is a call that feels broken to one caller in ten.
  • Per-call breakdown with a bar for every answer, so a single stall is visible instead of averaged away.
HOW LONG CALLERS WAITTo hear “hello”1.45s2.10susuallyon a slow callTo get an answerone bar per answer
KNOWLEDGE BASE

Your knowledge base, fact-checked against real questions

Crawling a website gets you whatever the business chose to publish, which is rarely what callers ask. Nutaan now puts the real questions to your knowledge base, reads what comes back, and tells you only what is genuinely missing.

  • Judged by reading the retrieved passages, not by a similarity score — a menu page scores 0.44 against “products and services” and used to be reported missing.
  • Asks in your trade’s words: a cafe is asked about its menu, a clinic about treatments, a dealership about models.
  • A third verdict for things half-there — items listed with no prices, a booking form with no rules.
  • Shows the passages behind every verdict, so “we don’t have your prices” is checkable rather than asserted.
WHAT YOUR AGENT CAN ANSWERMenu items and pricescoveredOpening hourscoveredReservation rulespartly thereWho to hand a hot lead tomissing
CONVERSATION

“Uh-huh” no longer interrupts the agent

People say “yeah”, “right”, “haan ji” while they are listening, not to take the floor. The agent now tells the difference by what was said rather than how many words it was — and a plain “yes” after a question still counts as the answer.

  • 78 acknowledgements shipped across English, Hindi and Hinglish, applied by default.
  • Add your own per agent, for a language we don’t ship yet.
  • Judged by content, not word count: “haan ji haan ji” is four words and doesn’t interrupt; “no wait” is two and does.
WHO GETS TO INTERRUPT“uh-huh”, “haan ji”keeps talking →“ok but what’s the price?”stops instantly
CALLING

Dial by hand, from your own number, without leaving

Not every call should be automated, and the ones that shouldn’t were forcing your team into a second tool. The dialer is now part of the same workspace, using the same number, writing to the same history.

  • Call straight from the browser on your configured Nutaan number.
  • Recording, transcript and outcome land in the same place as the AI calls.
  • No second subscription and no context lost between platforms.
nutaan.com/dialer123456789*0#Live callyour own number00:27 · recordingHang up
BENCHMARKS

Official benchmarks, published

Everyone in voice AI claims low latency and high accuracy. We published the measurements instead — per language, per engine, with the method written down so you can repeat them.

  • Speech recognition accuracy by language, including Indian English and regional languages.
  • Engine-by-engine response times, measured on real calls rather than a lab loop.
  • The method is stated, so the numbers can be checked rather than believed.
MEASURED, NOT CLAIMEDEnglish (US)96%English (IN)92%Hindi88%Bengali81%Marathi78%
LANGUAGES

Agents that follow the caller’s language mid-sentence

Real Indian calls switch between Hindi and English inside one sentence, and a US caller expects plain English throughout. An agent now follows whoever it is talking to instead of dragging them back.

  • Answers in the language the caller used, switching whenever they do.
  • Technical and commercial terms stay in English, which is how people actually say them.
  • 30+ languages, with recognition tuned per language rather than one global setting.
IT FOLLOWS THE CALLER“What’s the price?”“It starts at ₹4 a minute.”“हाँ जी, थोड़ा detail बताइए”

Every one of these is live. Hear it on a real call.

Point it at your website, and it answers the way the entries above describe — measured, in your caller’s language, without cutting them off.