Voice, pictures and language
Customer voice notes are turned into text on your own computer and answered like any other message; if you choose, replies go as voice notes in your own voice. Your photos and files are described and sent from one library, and every customer is answered in Persian, Arabic or English.
Recorded from the real app, with a sample business's data.
A customer's voice note is turned into text on this computer, offline, and answered like a written message: from the answer bank, the flows or the AI.
The language of a voice note is taken from the customer's earlier messages or detected. When the detection is unsure, or another language is heard, the business's main language is used.
If a voice note is not clear, the customer is asked in their own language to send it again or type it, and what was heard is noted on the customer's file for the team.
The inbox shows that a voice note is being read, then its text with the note's length; the text is also in search and in the conversation history.
A long voice note is cut at the speaker's pauses into pieces of about 20 seconds, each read on its own, so no sentence is lost.
Your brand, service names and catalogue product names are given to the reader as hints, so they are written the way your business writes them.
WhatsApp and Telegram voice notes, Instagram voice notes and the usual audio files are all read.
The app picks the speech model by your computer's power, and the owner can change it. Only the accurate model reads customers' notes; on a weak computer a voice note gets the 'file received' reply and goes to the team.
On the System page, try a sample voice note in each language: you see what was heard, what was really said and how long it took.
Voice notes are read apart from the rest of the work, so other customers do not wait, and notes left half-read by a restart are read again.
Up to 15 minutes of each person's voice notes are read per day, so one person cannot tie up the computer; beyond that the usual 'file received' reply goes.
A customer's audio file is read on its own only, with no network or other file access and with a time limit.
When a customer sends a voice note, the AI keeps its answer shorter, and the customer's file gets a 'Sends voice notes' tag the team can sort and group by.
In the simulator, record a voice note with your microphone or send an audio file and see how it is read and answered, with no real customer involved.
When switched on, the bot's reply goes as a voice note in a voice made from the owner's own. The full reply text stays in the panel.
You can answer with voice only the customers whose last message was a voice note, while everyone else gets text.
Before the voice is made, the reply is rewritten in spoken style: colloquial Persian, spoken Gulf Arabic, casual English; prices, numbers, times and product names stay locked.
The voice is made on this computer and no data leaves it, in all three languages. The model file is downloaded only when you press the button and is checked before use.
A lighter local engine for Arabic and English that works with 8 GB of RAM; it is tried automatically after your chosen engines when it speaks the reply's language.
Your voice is made with your own ElevenLabs account. A 30-second reply costs about 3–4 cents and counts against your AI budget.
Your voice is made with your own paid Gemini key. A 30-second reply costs under one cent. One more short recording is needed: Google's consent sentence in English or Arabic.
On WhatsApp and Telegram a real voice message with a play button goes out, on Instagram an audio file. The voice note is kept under WhatsApp's size limit.
Under every voice note comes one written line saying the voice was made with AI from your voice; you set the name in it. The reply's buttons come with this line.
Replies the bot has sent word for word again and again are made and checked ahead in idle time, so next time they go out at once.
If a customer writes 'text please', «متن بفرست» or «نص فقط», they get text from then on; asking for voice again brings voice back.
See how many voice notes went out in the last 30 days, how many replies went as text instead and why, and the cloud cost.
Every number, price, time, date, phone number and percentage is written out in words by the app's fixed rules, not by the AI, in everyday forms.
With a cloud AI connected, a separate check asks whether the meaning is the same; if not, the reply's original words are spoken.
If the colloquial rewrite is refused, the reply can go as text instead of reading out the formal words.
If the main engine fails, the other engines are tried in the order you set, and in the end the same reply goes as text.
Switch voice replies on or off separately for Instagram, WhatsApp and Telegram.
While the voice is being made, Telegram shows the customer a 'sending a voice' status and Instagram shows typing.
Beside the time limit you see roughly how long a 10-second voice note takes to make on this computer.
The made voice is turned back into text and compared with the reply: every number in order, every 'not', and any extra sentences. If anything differs, it is made once more, otherwise the same reply goes as text.
Sensitive replies always go as text, so the customer can read them exactly and keep them.
If the voice is not ready within the time you set, the same reply goes as text, and a late voice is never sent after the text.
Whatever goes wrong in making or sending the voice, the same reply is sent as text.
A voice reply is made only when at least 10 minutes are left in the messenger's 24-hour reply window.
The local engine starts only with enough free memory and stops at once if memory runs low; that reply goes as text.
The cost of cloud voice engines counts in the monthly and daily AI budget; when it is spent, the local engine or a text reply takes over.
The recording, its text, the samples, the ready voice notes and the copies made at cloud services are all deleted, and voice replies switch off.
Before any recording is kept, the owner confirms the voice is their own and agrees to its use; the time, language and engine of that consent are recorded.
The owner reads the shown text into their microphone in the panel; a level meter and seconds are shown, and poor recordings get a warning.
The recording and its text are kept encrypted on this computer and opened only for the moment a voice is made.
Three short samples in your voice are made and checked, including a number and a time; no reply goes as voice until you approve them.
What is heard in the recording is compared with the shown text; if they differ a lot, the owner corrects the text to what was really said or records again.
Voice reply settings and the owner's voice are open to the owner only; staff cannot even view them.
Before a cloud engine is first used, the service's data notice and country rules are shown and the owner confirms reading them.
Upload your price list, intro video, PDF, voice note and photos once; flows, automations, the inbox and the AI send them to customers.
Every library photo without labels is looked at by the picture AI: a description in Persian, Arabic and English, a few labels, and the text written on the photo.
Library files are sent on Instagram, WhatsApp and Telegram. On Instagram, after the first send the file stays ready for fast sending.
The AI knows what each file shows from its title and labels, and sends exactly the file the customer asked for.
The owner can edit any file's description and labels; the owner's labels are never replaced by the AI's, and when a file is replaced its old AI labels are cleared.
For a product without a photo, the smart setup suggests a library photo whose labels and description name that product. Photos you set yourself are never replaced.
Shop and catalogue
Preview photos, videos, audio and PDFs, rename, copy the key, delete, and replace a file while keeping its key; each file shows how many times it was sent.
Catalogue product photos, listing photos, reservation and rental photos, exercise videos and industry pack files are all picked from this library, and a new file can be uploaded right from that page.
Shop and catalogueListingsStays, tables and rentalsMemberships and subscriptions
Photos are seen by a built-in picture model on this computer, or by the cloud AI or an Ollama model that can see. The picture engine runs only while working and stops after a few idle minutes.
Library files are made available to the messenger through an unguessable link for sending; customers' personal files never go into the library and are never public.
Each customer's language is detected from their messages and kept on their file; replies, reminders and scheduled messages go in that language.
Persian typed in Latin letters is recognised as Persian, and Arabic typed in Latin letters and digits as Arabic; the answer goes in the right language.
A text you wrote in one language only, such as a service name, the welcome or a flow message, is translated into the customer's language just before it goes; button titles too.
Short messages like 'ok', «ممنون» or an emoji never change a customer's language; a switch needs a confident detection.
If a number, price or link goes missing or changes in a translation, its length looks wrong, or it is not in the right language, the original text is sent.
Set which of the three languages your business answers in and which one is the default; the brand name can be written separately in each language.
The health check before publishing shows which saved message, form or flow has no text in one of your business languages, so you can fill it in before a customer sees it.
Persian and Arabic share one script; they are told apart by letters only one of them uses, frequent words and the kind of digits.
Translations are written in a warm chat tone: Persian with the polite «شما», Gulf Arabic, casual English; emojis are kept and service names are translated.
Your business address and your own words, such as the staff title and what you call a customer, an offering, a booking and an order, are set per language, and the bot uses those words.
The whole admin panel is in three languages, right-to-left for Persian and Arabic; the panel language is picked at the first setup step and can be changed later.
The bot's ready messages, customer pages and forms come in three languages; a text not yet written in Arabic is shown to an Arabic reader in English first, not in Persian.
Persian shows Persian digits and the Persian (Jalali) calendar; Arabic and English use the digits common in the Gulf; the Umm al-Qura Hijri calendar is supported too.
More of the product
Hi! Ask me anything about DMAI Pro: features, pricing, installing, security.
