AI · Conversational, Voice & Multilingual AI
The interface is the conversation. The system behind it is the point.
Noble builds conversational and voice interfaces that sit over an organisation’s real systems — the service desk, the account, the order, the record — so that a question in natural language, typed or spoken, in Arabic or English or both in one sentence, gets an answer from the system that actually knows, and a request becomes an action with a record. The language layer is designed for the way people in this region actually speak and write, which is not the way a translated English design assumes.
Arabic first, and what that actually means
Most conversational systems in this region are English systems with Arabic added: the intents were designed in English, the test set was written in English, and the Arabic was translated in afterwards. They fail in predictable ways — on dialect, on a customer who writes Arabic in Latin letters, on the sentence that switches to English for the product name and back, on numbers, dates and names that are written differently in each script. Noble designs the language layer from the Arabic side: the evaluation set is written by Arabic speakers in the register customers actually use, the interface is composed right to left rather than mirrored, and English is the second language the system is equally good in.
Voice adds a layer with its own failure modes. Transcription has to hold on a noisy shop floor or a phone line, in a dialect, with a product code spelled out letter by letter. Noble uses voice where it earns its place — where the hands are busy, where the eyes are on something else, where the person is on a call already — and builds the confirmation step in, because a spoken order that was misheard is a mistake the organisation pays for.
The stack between a person and the system
Read from the top: the channel the person is on, the speech layer if they are speaking, the language layer that understands Arabic and English as they are actually used, the grounding that reaches the real system, and the system of record itself.
- 01Channel — chat, messaging, web, phone, in-appWhere the person already is
- 02Speech — transcription and synthesisDialect, noise, spelled codes
- 03Language — Arabic and English, mixedIntent, entities, register
- 04Grounding — retrieval, APIs, actionsOn the person’s permissions
- 05System of recordService desk, account, order
What the service covers
Conversational interfaces over real systems
Not a scripted tree, and not a chat window with nothing behind it: an interface that reads the account, checks the order, opens the case and books the visit through the systems that own those things, on the person’s own permissions. Designed with a clear boundary for what it can do alone, what it confirms first, and when it hands to a person with the conversation attached.
- Reads and acts through the systems of record
- On the person’s own identity and permissions
- Confirms before it commits; hands off with context
- Customer, employee and field use cases
Arabic and multilingual handling as a first concern
Dialect and Modern Standard Arabic both understood; Arabic written in Latin letters recognised; code-switching to English mid-sentence handled as the normal case; numerals, dates and names normalised across scripts. The evaluation set is written by Arabic speakers in the register the organisation’s customers actually use, and the interface is composed right to left rather than mirrored from an English layout.
- Dialect and Modern Standard Arabic
- Code-switching and Latin-script Arabic handled as normal
- Numerals, dates and names normalised across scripts
- Evaluation sets written in Arabic, not translated into it
Voice, transcription and spoken interaction
Speech in and speech out where it earns its place: the field engineer reporting a reading with both hands on the equipment, the caller who would rather say it than type it, the meeting that needs a transcript in the language it was held in. Transcription is evaluated on the organisation’s real audio — its noise, its dialect, its product codes — and a spoken request that will change a record is read back before it is committed.
- Transcription evaluated on the organisation’s real audio
- Spoken requests read back before they are committed
- Field, call-centre and meeting use cases
- Arabic and English synthesis composed for each
Natural language processing on text at volume
Classification, summarisation, entity extraction and routing over the text an organisation receives every day — emails, tickets, complaints, notes, messages — in both languages. Built to route and summarise rather than to decide, with the confidence carried and the low-confidence cases going to a person.
- Classification, summarisation and extraction in Arabic and English
- Routing of tickets, emails and complaints with confidence
- Summaries that cite the passages they came from
Enterprise search across permitted content
One search over the content people are allowed to see — documents, records, tickets, knowledge — with results filtered by the same permissions as the sources, ranked for the question rather than the keyword, and answered directly where the answer is in the content. Arabic and English indexed together, so a question in one finds a document in the other.
- Permission-filtered results from the same identity system
- Ranked for the question, answered where the content allows
- Cross-lingual retrieval between Arabic and English
Consent, identity and the record
A conversation is personal data. Who the person is, what they consented to, what the system retrieved and what it did are recorded against the identity, kept for the period the organisation’s rules require, and available to the person who asks what was done in their name. Voice recordings are held under a written retention rule decided before the first call.
- Consent and identity established before anything is retrieved
- Every retrieval and action logged to the person
- Retention for transcripts and recordings decided in writing
Where this connects
- Generative AI & LLM ApplicationsThe grounding behind a conversation — retrieval on the person’s permissions, answers that cite their sources — is designed there.
- AI Agents & Intelligent AutomationWhen a conversation has to do something — book, change, cancel — the action runs under the permission and stopping rules on that page.
- AI for Enterprise SystemsThe systems of record at the bottom of the stack — and how they are exposed safely to a conversation — are that page.
The services alongside
- Software & Web EngineeringThe channel itself — the web widget, the mobile screen, the messaging integration, the right-to-left interface — is built by this service.
- SecurityIdentity, consent and the permissions a conversation runs on come from the same system as everything else.
- Managed IT ServicesWhere the voice channel is a phone system, the telephony beneath it is supplied and supported through Managed IT Services.
Bring a hundred real messages.
As customers actually wrote them — dialect, Latin-script Arabic, the English product name mid-sentence. That, with what the right response was, is the evaluation set, and it says more than any demonstration.