Arabic voice notes in one bot: qualification that works

A single WhatsApp bot can qualify a Dubai real estate lead in Khaleeji Arabic, English, or a mix of both. The pipeline is the same either way: transcribe each voice note, extract and confirm the qualifying fields in the lead's language, then score and route. The fragile part is transcription, not the bot

What does an Arabic voice-note qualification bot actually do?

The ask is a property enquiry. In Dubai that usually starts as a voice note sent to an agent's WhatsApp or a business number: the buyer speaks their budget, the area they want, the unit type, and sometimes a second property text in a different sentence. A text bot reads the enquiry directly. A voice-note bot has to get the same fields out of audio first

The pipeline we build for brokerages runs in six steps, and each one has a named output that a human can audit

1. Capture. The incoming voice note arrives on the WhatsApp Business Platform, transcript or not, and a webhook hands the audio to the transcription step. Meta delivers media to the business through the platform, so this works on the official API
2. Transcribe. A speech-to-text model converts the audio to text. For Gulf Arabic this is where quality varies most, because models trained on Modern Standard Arabic misread everyday Khaleeji speech, including the shift of qaf to /g/ or /j/ and Gulf-specific function words that rarely appear in MSA corpora. A good build uses a dialect-aware model and records a confidence score with the transcript
3. Normalize. The raw transcript is full of contractions, borrowed English property terms, and numbers spoken in Arabic or English. Normalize budget figures into AED, map "دبي مارينا" to "Dubai Marina", and collapse unit types ("studio", "مكتب", "office") into one field vocabulary before anything is scored
4. Confirm. The bot repeats the extracted budget, area, and unit type back to the lead in their language and asks for a yes or a fix. This step catches the transcription errors before they reach the scoring rules
5. Score. Confirmed fields feed the same scoring rules a text enquiry would use, based on budget fit, area, property class, and how ready the buyer seems from what they said
6. Route and log. The lead goes to the agent whose desk it fits, with the transcript, confidence score, and confirmed fields attached, and the conversation writes back to the CRM

The confirm step is the one most naive builds drop. Without it, a misheard budget of 900,000 AED gets scored as if the lead had genuinely asked for it, and an agent wastes a call discovering a 2,400,000 AED gap

Every voice-note build we ship confirms the extracted budget and area back to the lead before scoring. Misheard numbers are the main source of junk leads, and the confirmation step removes them without the lead noticing any friction.

Do I need separate Arabic and English bots?

No. One bot with per-message language detection works better than two, because Dubai enquiries routinely switch language inside a single thread, sometimes inside a single voice note. A buyer writes in Arabic, then replies to a price with an English "ok sounds good send viewing times", then sends a Khaleeji voice note. Two separate bots force the lead to pick a lane; one bot just detects the language on each message

The qualification questions, the CRM fields, and the scoring rules stay identical in both languages. Only the reply text and the transcription model change. Keeping one field map and one scoring layer in both languages is what makes the pipeline testable, because the same confirmed field set can be scored by rules that do not care how the lead expressed it

Two implementation details keep a single bot honest. Language detection should run on the message level, not once per thread, so a code-switching lead is answered in whatever they just said. And the transcription step needs separate handling for Arabic audio and English audio, because forcing an English-optimized model onto Khaleeji speech is where Word Error Rate climbs

What is the accuracy of Arabic speech-to-text on WhatsApp?

For clear, single-speaker Khaleeji, modern transcribers do well. The failures concentrate on the fields a property enquiry depends on: names, area spellings, and numbers. A model tuned for broadcast Arabic will give you a confident transcript of a Marina apartment enquiry that is wrong on three digits of the rent, because Gulf dialects differ from MSA in vocabulary, phonology, and grammar, and models trained on MSA misread everyday Khaleeji speech. Code-switching adds another failure mode, since a voice note mixes Arabic and English property terms in a way that plain MSA or plain English data does not prepare a model for

The practical fix is not chasing a perfect Word Error Rate. It is isolating the fields and confirming them, which is why the confirm step exists. Treat the transcript as draft data, confirm the budget and area back, and only then let a number into the scoring rules. A field that is confirmed by the lead is accurate regardless of what the base model heard

Which dialect failure modes actually cost you a lead?

The table below lists the transcription failures we see in real property voice notes and the mitigation that works for each. The modes are specific to Gulf Arabic; a bot built for Levantine or Egyptian speech will have a different set

Failure mode	What the model does	Cost if unchecked	Mitigation
qaf shifting to /g/ or /j/	Hears "qamar" as "gamar", misreads area names and landmarks	Area field wrong, lead routed to the wrong desk	Dialect-aware model plus confirm the area back in Arabic
English property terms inside Arabic speech	"studio", "Dubai Hills", "AED" borrowed mid-sentence	Half the transcript comes back as mixed-script noise	Separate Arabic and English transcription models per message
Numbers spoken in Arabic or English	Rent and budget words misheard by one or two digits	Scoring puts the lead in the wrong budget band	Normalize to AED, then confirm the figure back
Names misheard	Double letters and soft consonants collapsed	Agent calls the wrong name, looks unprepared	Never rely on the transcript for the lead's name; confirm at handoff
Area names brand+project tokens	"Emaar", "Falcon City", "DIFC" reduced to phonetics	CRM gets a fuzzy location that no agent can match	Map transcript against a project and community dictionary before scoring

Each row pairs a specific failure with a specific step in the pipeline. That is the point of writing the modes down: a defect in the transcript maps to one place you fix, not a vague "improve the AI" request. Area and community names fail differently from budget numbers, so they need different guards, an entity dictionary for one and a confirm question for the other

How do you route and hand off a voice-note lead?

The transcript and confirmed fields are the context you hand to a human, and they have to be readable in Arabic as well as English, because the agent who closes the sale may work either language. The handoff bundle is the message thread, the transcript, a one-line summary of what the lead wants, and the reason the bot is escalating

Escalation is a decision with a threshold, and we set it low on the audio channel. Text can be re-read, so a text bot can afford a longer unconfirmed stretch. A voice-note lead that was already misheard once should not accumulate more guesses. When the transcript confidence is low, when two confirmation attempts fail, or when the lead asks a question the qualification flow does not cover, the bot hands off with everything it has so far. The lead never has to restate the budget that already came out of their first voice note

Meta's Business Messaging Policy requires the customer to have a clear path to a human, and that applies to voice-note automation the same way it applies to a text bot. The handoff button is the difference between a tool that qualifies and a wall that frustrates. We cover the escalation rules and thresholds in detail in human handoff: when a WhatsApp bot must escalate

Where does the voice-note data live, and is that PDPL clean?

An audio recording of a buyer saying their budget is personal data under the UAE Personal Data Protection Law (Federal Decree-Law No. 45 of 2021), and the broker is the controller whether the transcription runs locally or on a hosted model. The safe pattern is to record the retention period for audio and transcripts, define which subprocessors see the audio during transcription, and give the lead a way to have the recording deleted. We walk through the full data-flow map for a CRM-and-WhatsApp stack in PDPL-compliant AI CRM in the UAE.

A practical rule is to keep the raw audio only as long as the transcript needs it. Once the confirmed fields are in the CRM and the agent has the handoff bundle, the recording adds retention risk with little operational value. The transcript persists as a thread record, which is normal; the audio is what most operators choose to expire

Does a bilingual voice bot work on your existing stack?

Yes, and it is designed not to replace your CRM. The bot writes the confirmed field map, the transcript, and the conversation thread into HubSpot, Salesforce, Zoho or Odoo through the same write-back the text intake uses, so one contact record holds the whole journey. The field map for a voice-note lead is the same schema a form submission or a portal enquiry fills, which is what makes the rest of your operation indifferent to how the enquiry arrived. Our WhatsApp + HubSpot, Zoho, Odoo integration write-up covers the field mapping and dedup

The assessment for a voice-note channel is the same one you would run on a text bot: does the lead volume justify the transcription spend, does the team have an agent who can close Arabic conversations, and is the handoff rule defined before launch. The reuse is the point. A brokerage that already qualifies Property Finder text leads on AI lead qualification adds a voice-note route by pointing the same field map and scoring rules at audio input

On the a property-management platform property management product we built as an MVP in 10 days, around 60% of tenant chats are handled by AI and the rest go to a person. The same split applies to a voice-note intake: the bot qualifies the straightforward enquiries and routes the ambiguous ones, and the handoff decision is made before launch, not improvised mid-conversation

The takeaway

An Arabic and English voice-note bot is a text-intake bot with a fragile transcription step bolted on the front. The engineering that makes it work is not a better speech model; it is the six-step pipeline where transcription feeds normalization, confirmation, and scoring, and where a confirmed budget never reaches the scoring rules unverified. Keep one field map in both languages, mirror the lead's language at every reply, expire the raw audio, and define the handoff threshold before you ship

If you want to know what a Khaleeji voice-note channel would cost on your lead volume, book the free audit and we will map your current property enquiries, check the transcription failure modes against your area and community names, and give you a fixed number for the first build