Voice notes, and three languages
Typing Hindi or Gujarati on a phone is slow enough that people simply do not do it. So you can talk instead, and a spoken instruction walks the identical confirmation steps a typed one does, because it becomes text and joins the same path.
Send a voice note and Robnu transcribes it, then hands the text to the same router every typed message uses. English, Hindi and Gujarati are all answered in kind, including romanised Hindi and Gujarati typed in Latin letters. Voice adds no abilities that typing does not have.
- A voice note is turned into text and handed to the same router a typed message goes to, so a spoken instruction confirms exactly as a typed one does.
- Voice is not a second assistant and grants no extra abilities, because two definitions of what a message can do would leave the spoken one weaker.
- Three languages are answered in kind: English, Hindi and Gujarati, in their own script.
- Romanised Hindi and Gujarati typed in Latin letters are covered too, which script detection alone cannot do.
- Robnu does not guess beyond those three, because a script can carry several languages and a confident wrong guess once answered a Gujarati seller in Urdu.
Two things make a chat usable for sellers who do not work in English. The first is obvious: answer in their language. The second is less so: answer in the same language consistently, so a Gujarati question does not get a Gujarati answer followed by an English menu. Getting that right turned out to need two mechanisms rather than one, because the easy half of the problem is the half that detection can see.
Why voice, and not just better typing
Typing Hindi or Gujarati on a phone keyboard is slow enough that people do not do it. They hold the button and talk. That is not a preference to accommodate, it is how a large share of Indian sellers already use every other app on the phone.
Before this worked, a voice note arrived, was downloaded, and then nothing happened. No reply, no acknowledgement, no record that anything had been said. The seller had no way to tell whether Robnu was broken or ignoring them, which is worse than a plain refusal.
What was missing was never a model. The downloader, the transcriber and even the column that stores the transcript all already existed. The gap was the wire between them, which is worth saying because it is the usual shape of a feature that looks absent.
A voice note becomes text, then takes the ordinary path
The audio is transcribed and the resulting text is handed to exactly the same router every typed message goes through. The same keyword commands, the same processing flow, the same read only assistant, the same permissions.
So a spoken request to process orders walks the identical confirmation buttons a typed one does. It does not shortcut them, and it cannot, because by the time the router sees it there is nothing to distinguish it from text.
This is deliberate rather than convenient. Voice becoming its own assistant would mean two definitions of what a message is allowed to do, and the spoken definition would be the weaker one, because it is the one built later and tested less. One path means one set of guarantees.
Three languages, answered in their own script
English, Hindi and Gujarati. A message in Devanagari or Gujarati script is identified by the script itself, which is unambiguous, and the canned replies and menus come back in that language.
Consistency is the point. A seller who asks in Gujarati, gets an answer in Gujarati, and then receives a menu in English has been given a worse experience than one answered entirely in English, because the mixed reply reads as a system that half works.
That was a real defect rather than a hypothetical. A typed Gujarati question fell back to English because nothing in the message told the system what it was, and the sellers it affected are not an edge case.
The hard half: Hindi and Gujarati in Latin letters
Script detection solves the easy half. It cannot touch the common case where someone types Hindi or Gujarati in Latin letters, because a romanised Gujarati sentence is genuinely indistinguishable from English by character alone.
That half is covered differently: the assistant is held to answering in the same language and script the seller used, whatever those turn out to be. So a romanised question comes back in kind without anything having had to classify it first.
Two mechanisms for one goal sounds redundant and is not. Detection is certain where it applies and blind elsewhere. The standing instruction is less precise and has no blind spot. Each covers what the other cannot.
Hold the button and say it
Transcribed into text, then treated exactly as typing
What Robnu deliberately will not guess
A script is not a language. Devanagari carries Hindi, Marathi, Nepali and more. Arabic script carries Urdu, Arabic and others. Guessing the language from the script is therefore a confident guess that is sometimes simply wrong.
It has been wrong here before, in a way worth remembering: a guess of that kind once answered a Gujarati seller in Urdu. That is not a small cosmetic error, it is a reply the person cannot read, produced with complete confidence.
So the canned replies cover three languages and stop. A message in a script outside those is still understood and answered by the assistant, mirroring whatever the seller wrote, rather than being forced into a language a lookup table picked for them.
The part of this Robnu does for you
Robnu reads your Amazon account and reconciles every settlement line back to the order that caused it, so a wrong deduction is visible the day it lands.
- Orders, returns, payouts and inventory synced
- Every settlement line matched to its order
- FBA units Amazon has not made good, listed
Where this sits in the rest of your Amazon operation
On Amazon, Robnu syncs your orders, returns, payouts and inventory, reconciles every settlement line back to the order that caused it, lists the FBA units Amazon's own ledger has not made good, flags a return that never came back before the SAFE-T window closes, and scans returns in at your dock. It reads your Amazon account rather than writing to it, so accepting orders and making labels run on AJIO and Meesho today. Related reading:
Sources & further reading
- Robnu, the product the behaviour described here is the product's own. The WhatsApp surface calls the same endpoints the screens do, with the same permissions, so anything you can check on a screen is the same fact. Read 3 November 2026.
- Meta, WhatsApp Business Platform onboarding Meta's own page. It states that the 24 hour customer service window applies to messages sent through the Cloud API, and that a window opens when a person messages a business already onboarded to it. That rule is why a reply can arrive as a template rather than as free text. Read 3 November 2026.
Not legal or tax advice. Amazon changes fees, windows and policies without notice, and the detailed style guides and account-health thresholds sit behind a Seller Central login, so any figure quoted here as convention is not Amazon-published. Check anything you plan to act on against your own Seller Central account first.
Questions sellers ask
Yes, and it walks exactly the same confirmation steps as typing. The voice note is transcribed and the text is handed to the same router a typed message uses, so by the time anything acts on it there is nothing to distinguish the two. Voice grants no abilities that typing does not already have.
English, Hindi and Gujarati, each in its own script. Messages in Devanagari or Gujarati script are recognised by the script, which is unambiguous, and the replies and menus come back in that language rather than switching to English partway through.
That is covered, by a different mechanism. Script detection cannot tell romanised Hindi or Gujarati from English, because by character they are the same. Instead the assistant answers in the same language and script you used, so a romanised question comes back in kind without anything having to classify it first.
Because a script is not a language, and guessing wrongly is worse than not guessing. Devanagari carries Hindi, Marathi and Nepali among others. A guess of that kind once answered a Gujarati seller in Urdu, which is a reply the person cannot read at all, delivered with full confidence. Three canned languages plus an assistant that mirrors whatever you wrote avoids that.
No, and that is the important part. It is transcribed into text and joins the ordinary path, so there is one definition of what a message can do. A separate voice assistant would be a second definition, and being newer and less exercised it would be the weaker of the two.
Yes, images are read as well, and they follow the same principle as voice: the content becomes something the ordinary router can handle, rather than opening a separate path with its own rules and its own permissions.
Related seller guides
More on the operations, money and claims that decide whether a marketplace catalogue actually makes money.
Amazon bullet points, how many and how long each should be
Amazon's public page states no bullet count and no character limit at all. What it does say, what the category guides hold, and how to write bullets that answer the objection instead of listing features.
Amazon Reconciliation: What It Is, And What It Is Not
Reconciliation is not bookkeeping. It is proving that what Amazon paid you matches what Amazon owed you, line by line, and finding the gap when it does not. Here is the method, the eight places money goes missing, and what to automate.
Receive Returns By Sending Codes In A WhatsApp Chat
Say return scan, then send AWB after AWB as the parcels come off the van. Each one is marked received the moment it lands, and shows up everywhere a scan from the web dock does, reports included.
Capture A Claim On WhatsApp, With Photos And Video
Say start claim, describe what happened in any language, and send the AWB photo, the barcode, a photo of what you received and the opening video. Robnu reads each photo and tells you what it read. The draft is finished and submitted from the app.
Every Robnu WhatsApp Command, And What Each One Does
The short commands Robnu matches exactly, the modes they open, and the much larger set of things you can simply ask in a sentence. Including which ones change something and which only read.
Amazon Settlement Report: Mapping Old Columns To New
Amazon's changelog does publish a mapping from the old settlement columns to the new three. Here is each pairing, the consolidation that makes totals move, and how to reconcile the two reports over one period before 11 November 2026.
Process Your Marketplace Orders From A WhatsApp Chat
Say process my orders in a WhatsApp chat, pick the account, confirm, and Robnu runs the real accept, label and manifest pipeline. Four gates stand between a tap and a live run, and the assistant is not one of them.
Download A Shipping Label From WhatsApp In One Message
Send an order number, an AWB or a time window and the label PDF arrives in the chat, built to your own label settings. This path only re-downloads a label you already own. It never accepts an order and never mints a new one.